Gateway to think about things in demand .: Sybase Adaptive Server Enterprise

Adaptive Server Enterprise

0. ASE History
0.1 What's in a name?
0.2 A History of ASE versions
1. ASE Administration
1.1 Basic Administration
1.2 User Database Administration
1.3 Advanced Administration
1.4 General Troubleshooting
1.5 Performance and Tuning
1.6 Server Monitoring
2. Platform Specific Issues
2.1 Solaris
2.2 NT
2.3 Linux
3. DBCC's
4. isql5. bcp
6. SQL Development
6.1 SQL Fundamentals
6.2 SQL Advanced
6.3 Useful SQL Tricks
7. Open Client
9. Freeware
10. Sybase Technical News
11. Additional Information
12. Miscellany

0.1 What's in a name?

Throughout this FAQ you will find references to SQL Server and ASE or Adaptive Server Enterprise to give it its full name. Do not be confused by the fact that Microsoft also has a product called SQL Server.

Well, back at about release 4.2 of Sybase SQL Server, the products were exactly the same. Microsoft were to do the port to NT. Well, it is pretty well documented, but there was a falling out. Both companies kept the same name for their data servers and confusion began to reign. In an attempt to try and sort this out, Sybase renamed their product Adaptive Server Enterprise (ASE) starting with version 11.5.

The following is a quote from a Sybase manual:

Since changing the name of Sybase SQL Server to Adaptive Server Enterprise, Sybase uses the names Adaptive Server and Adaptive Server Enterprise to refer collectively to all supported versions of the Sybase SQL Server and Adaptive Server Enterprise. Version-specific references to Adaptive Server or SQL Server include version numbers.

We will endeavour to try and do the same within the FAQ, but the job is far from complete!

0.2 A History of ASE Versions

The following is a list of the versions of ASE released over the years:

· Sybase SQL Server 3.0 – the first publicly released version of the product released in 1988.

· Sybase SQL Server 4.0 – released in 1990.

· Sybase SQL Server 4.2 – released in 1991.

· Sybase SQL Server 4.8 – released in 1992.

· Sybase SQL Server 4.9 – also released in 1992, along with version 4.9.1; version 4.9.2 was released in 1993.

· Sybase System X, aka Sybase SQL Server 10.0 – released in 1993.

· Sybase SQL Server 11.0 – released in 1995.

· Adaptive Server Enterprise 11.5 – released in 1997.

· Adaptive Server Enterprise 11.9 – released in 1998. It does not appear that version 11.9.0 was ever publicly released; instead 11.9.2 was the released version, and 11.9.3 was the first 64 bit version made available.

· Adaptive Server Enterprise 12.0 – released in 1999.

· Adaptive Server Enterprise 12.5 – released in 2001.

· Adaptive Server Enterprise 12.5.0.1 – released in 2002, along with version 12.5.0.2.

· Adaptive Server Enterprise 12.5.0.3 – released in 2003, along with version 12.5.1.

· Adaptive Server Enterprise 12.5.2 – released in 2004, along with version 12.5.3.

· Adaptive Server Enterprise 12.5.3a – released in 2004/2005.

· Adaptive Server Enterprise 12.5.4 – released in 2006.

· Adaptive Server Enterprise 15.0 – released in 2005, along with version 12.5.3.

· Adaptive Server Enterprise 15.0.1 – released in 2006.

· Adaptive Server Enterprise 15.0.2 – due for release in 2007.

1.1.1 What is SQL Server and ASE anyway?
1.1.2 How do I start/stop ASE when the CPU reboots?
1.1.3 How do I move tempdb off of the master device?
1.1.4 How do I correct timeslice -201?
1.1.5 The how's and why's on becoming a Certified Sybase Professional DBA (CSPDBA)?
1.1.6 RAID and Sybase
1.1.7 How to swap a db device with another
1.1.8 Server naming and renaming
1.1.9 How do I interpret the tli strings in the interface file?
1.1.10 How can I tell the datetime my Server started?
1.1.11 Raw partitions or regular files?
1.1.12 Is Sybase Y2K (Y2000) compliant?
1.1.13 How Can I Run the ASE Upgrade Manually?
1.1.14 We have lost the sa password, what can we do?
1.1.15 How do I set a password to be null?
1.1.16 Does Sybase support Row Level Locking?
1.1.17 What platforms does ASE run on?

1.4.1 How do I turn off marked suspect on my database?

Say one of your database is marked suspect as the SQL Server is coming up. Here are the steps to take to unset the flag.

Remember to fix the problem that caused the database to be marked suspect after switching the flag.

System 11

sp_configure "allow updates", 1

select status - 320 from sysdatabases where dbid = db_id("my_hosed_db") -- save this value.

begin transaction

update sysdatabases set status = -32768 where dbid = db_id("my_hosed_db")

commit transaction

shutdown

startserver -f RUN_*

fix the problem that caused the database to be marked suspect

begin transaction

update sysdatabases set status = saved_value where dbid = db_id("my_hosed_db")

commit transaction

sp_configure "allow updates", 0

reconfigure

shutdown

startserver -f RUN_*

System 10

sp_configure "allow updates", 1

reconfigure with override

select status - 320 from sysdatabases where dbid = db_id("my_hosed_db") - save this value.

begin transaction

update sysdatabases set status = -32768 where dbid = db_id("my_hosed_db")

commit transaction

shutdown

startserver -f RUN_*

fix the problem that caused the database to be marked suspect

begin transaction

update sysdatabases set status = saved_value where dbid = db_id("my_hosed_db")

commit transaction

sp_configure "allow updates", 0

reconfigure

shutdown

startserver -f RUN_*

Pre System 10

sp_configure "allow updates", 1

reconfigure with override

select status - 320 from sysdatabases where dbid = db_id("my_hosed_db") - save this value.

begin transaction

update sysdatabases set status = -32767 where dbid = db_id("my_hosed_db")

commit transaction

you should be able to access the database for it to be cleared out. If not:

shutdown

startserver -f RUN_*

fix the problem that caused the database to be marked suspect

begin transaction

update sysdatabases set status = saved_value where dbid = db_id("my_hosed_db")

commit transaction

sp_configure "allow updates", 0

reconfigure

Return to top

1.4.2 On startup, the transaction log of a database has filled and recovery has suspended, what can I do?

You might find the following in the error log:

00:00000:00001:2000/01/04 07:43:42.68 server Can't allocate space for object 'syslogs' in database 'DBbad' because 'logsegment' segment is full/has no free extents. If you ran out of space in syslogs, dump the transaction log. Otherwise, use ALTER DATABASE or sp_extendsegment to increase size of the segment.
00:00000:00001:2000/01/04 07:43:42.68 server Error: 3475, Severity: 21, State: 7
00:00000:00001:2000/01/04 07:43:42.68 server There is no space available in SYSLOGS for process 1 to log a record for which space has been reserved. This process will retry at intervals of one minute. The internal error number is -4.

which can prevent ASE from starting properly. A neat solution from Sean Kiely (sean.kiely@sybase.com) of Sybase Technical Support, that works if the database has any "data only" segments. Obviously this method does not apply to the master database. The Sybase Trouble Shooting Guide has very good coverage of recovering the master database.

You will have to bring the server up with trace flag 3608 to prevent the recovery of the user databases.

sp_configure "allow updates",1
go

Write down the segmap entries from the sysusages table for the toasted database.

update sysusages
set segmap = 7
where dbid = db_id("my_toasted_db")
and segmap = 3

select status - 320
from sysdatabases
where dbid = db_id("my_toasted_db") -- save this value.
go
begin transaction
update sysdatabases set status = -32768 where dbid = db_id("my_toasted_db")
go -- if all is OK, then...
commit transaction
go
shutdown
go

Restart the server without the trace flag. With luck it should now have enough space to recover. If it doesn't, you are in deeper trouble than before, you do have a good, recent backup don't you?

dump database my_toasted_db with truncate_only
go

Reset the segmap entries in sysusages to be those as saved in 3. above.

Shutdown ASE and restart. (The traceflag should have gone at step 6., but ensure that it is not there!)

Return to top

1.4.3: Why do my page locks not get escalated to a table lock after 200 locks?

Several reasons why this may be happening.

· Are you doing the updates from within a cursor?

The lock promotion only happens if you are attempting to take out 200 locks in a single operation ie a single insert, update or delete. If you continually loop over a table using a cursor, locking one row at time, the lock promotion never fires. Either use an explicit mechanism to lock the whole table, if that is required, or remove the cursor replacing it with an appropriate join.

· A single operation is failing to escalate?

Even if you are performing a single insert, update or delete, Sybase only attempts to lock the whole table when the lock escalation point is reached. If this attempt fails because there is another lock which prevents the escalation, the attempt is aborted and individual page locking continues.

Return to top

1.5.1: Sybase ASE Performance and Tuning

Before going any further, Eric Miner (eric.miner@sybase.com) has made available two presentations that he made at Techwave 1999. The first covers the use of optdiag. The second covers features in the way the optimiser works in ASE 11.9.2 and 12. These are Powerpoint slides converted to web pages, so they might be tricky to read with a text based browser!

All Components Affect Response Time & Throughput

We often think that high performance is defined as a fast data server, but the picture is not that simple. Performance is determined by all these factors:

· The client application itself:

· How efficiently is it written?

· We will return to this later, when we look at application tuning.

· The client-side library:

· What facilities does it make available to the application?

· How easy are they to use?

· The network:

· How efficiently is it used by the client/server connection?

· The DBMS:

· How effectively can it use the hardware?

· What facilities does it supply to help build efficient fast applications?

· The size of the database:

· How long does it take to dump the database?

· How long to recreate it after a media failure?

Unlike some products which aim at performance on paper, Sybase aims at solving the multi-dimensional problem of delivering high performance for real applications.

OBJECTIVES

To gain an overview of important considerations and alternatives for the design, development, and implementation of high performance systems in the Sybase client/server environment. The issues we will address are:

· Client Application and API Issues

· Physical Database Design Issues

· Networking Issues

· Operating System Configuration Issues

· Hardware Configuration Issues

· ASE Configuration Issues

Client Application and Physical Database Design design decisions will account for over 80% of your system's "tuneable" performance so ... plan your project resources accordingly !

It is highly recommended that every project include individuals who have taken Sybase Education's Performance and Tuning course. This 5-day course provides the hands-on experience essential for success.

Client Application Issues

· Tuning Transact-SQL Queries

· Locking and Concurrency

· ANSI Changes Affecting Concurrency

· Application Deadlocking

· Optimizing Cursors in v10

· Special Issues for Batch Applications

· Asynchronous Queries

· Generating Sequential Numbers

· Other Application Issues

Tuning Transact-SQL Queries

· Learn the Strengths and Weaknesses of the Optimizer

· One of the largest factors determining performance is TSQL! Test not only for efficient plans but also semantic correctness.

· Optimizer will cost every permutation of accesses for queries involving 4 tables or less. Joins of more than 4 tables are "planned" 4-tables at a time (as listed in the FROM clause) so not all permutations are evaluated. You can influence the plans for these large joins by the order of tables in the FROM clause.

· Avoid the following, if possible:

· What are SARGS?

This is short for search arguments. A search argument is essentially a constant value such as:

· "My company name"

· 3448

but not:

· 344 + 88

· like "%what you want%"

· Mathematical Manipulation of SARGs

SELECT name FROM employee WHERE salary * 12 > 100000

· Use of Incompatible Datatypes Between Column and its SARG

Float & Int, Char & Varchar, Binary & Varbinary are Incompatible;

Int & Intn (allow nulls) OK

· Use of multiple "OR" Statements - especially on different columns in same table. If any portion of the OR clause requires a table scan, it will! OR Strategy requires additional cost of creating and sorting a work table.

· Not using the leading portion of the index (unless the query is completely covered)

· Substituting "OR" with "IN (value1, value2, ... valueN) Optimizer automatically converts this to an "OR"

· Use of Non-Equal Expressions (!=) in WHERE Clause.

· Use Tools to Evaluate and Tune Important/Problem Queries

· Use the "set showplan on" command to see the plan chosen as "most efficient" by optimizer. Run all queries through during development and testing to ensure accurate access model and known performance. Information comes through the Error Handler of a DB-Library application.

· Use the "dbcc traceon(3604, 302, 310)" command to see each alternative plan evaluated by the optimizer. Generally, this is only necessary to understand why the optimizer won't give you the plan you want or need (or think you need)!

· Use the "set statistics io on" command to see the number of logical and physical i/o's for a query. Scrutinize those queries with high logical i/o's.

· Use the "set statistics time on" command to see the amount of time (elapsed, execution, parse and compile) a query takes to run.

· If the optimizer turns out to be a "pessimizer", use the "set forceplan on" command to change join order to be the order of the tables in the FROM clause.

· If the optimizer refuses to select the proper index for a table, you can force it by adding the index id in parentheses after the table name in the FROM clause.

SELECT * FROM orders(2), order_detail(1) WHERE ...

This may cause portability issues should index id's vary/change by site !

Locking and Concurrency

· The Optimizer Decides on Lock Type and Granularity

· Decisions on lock type (share, exclusive, or update) and granularity (page or table) are made during optimization so make sure your updates and deletes don't scan the table !

· Exclusive Locks are Only Released Upon Commit or Rollback

· Lock Contention can have a large impact on both throughput and response time if not considered both in the application and database design !

· Keep transactions as small and short as possible to minimize blocking. Consider alternatives to "mass" updates and deletes such as a v10.0 cursor in a stored procedure which frequently commits.

· Never include any "user interaction" in the middle of transactions.

· Shared Locks Generally Released After Page is Read

· Share locks "roll" through result set for concurrency. Only "HOLDLOCK" or "Isolation Level 3" retain share locks until commit or rollback. Remember also that HOLDLOCK is for read-consistency. It doesn't block other readers !

· Use optimistic locking techniques such as timestamps and the tsequal() function to check for updates to a row since it was read (rather than holdlock)

ANSI Changes Affecting Concurrency

· Chained Transactions Risk Concurrency if Behavior not Understood

· Sybase defaults each DML statement to its own transaction if not specified ;

· ANSI automatically begins a transaction with any SELECT, FETCH, OPEN, INSERT, UPDATE, or DELETE statement ;

· If Chained Transaction must be used, extreme care must be taken to ensure locks aren't left held by applications unaware they are within a transaction! This is especially crucial if running at Level 3 Isolation

· Lock at the Level of Isolation Required by the Query

· Read Consistency is NOT a requirement of every query.

· Choose level 3 only when the business model requires it

· Running at Level 1 but selectively applying HOLDLOCKs as needed is safest

· If you must run at Level 3, use the NOHOLDLOCK clause when you can !

· Beware of (and test) ANSI-compliant third-party applications for concurrency

Application Deadlocking

Prior to ASE 10 cursors, many developers simulated cursors by using two or more connections (dbproc's) and divided the processing between them. Often, this meant one connection had a SELECT open while "positioned" UPDATEs and DELETEs were issued on the other connection. The approach inevitably leads to the following problem:

Connection A holds a share lock on page X (remember "Rows Pending" on SQL Server leave a share lock on the "current" page).

Connection B requests an exclusive lock on the same page X and waits...

The APPLICATION waits for connection B to succeed before invoking whatever logic will remove the share lock (perhaps dbnextrow). Of course, that never happens ...

Since Connection A never requests a lock which Connection B holds, this is NOT a true server-side deadlock. It's really an "application" deadlock !

Design Alternatives

Buffer additional rows in the client that are "nonupdateable". This forces the shared lock onto a page on which the application will not request an exclusive lock.

Re-code these modules with CT-Library cursors (aka. server-side cursors). These cursors avoid this problem by disassociating command structures from connection structures.

Re-code these modules with DB-Library cursors (aka. client-side cursors). These cursors avoid this problem through buffering techniques and re-issuing of SELECTs. Because of the re-issuing of SELECTs, these cursors are not recommended for high transaction sites !

Optimizing Cursors with v10.0

· Always Declare Cursor's Intent (i.e. Read Only or Updateable)

· Allows for greater control over concurrency implications

· If not specified, ASE will decide for you and usually choose updateable

· Updateable cursors use UPDATE locks preventing other U or X locks

· Updateable cursors that include indexed columns in the update list may table scan

· SET Number of Rows for each FETCH

· Allows for greater Network Optimization over ANSI's 1- row fetch

· Rows fetched via Open Client cursors are transparently buffered in the client:

FETCH -> Open Client <- N rows

Buffers

· Keep Cursor Open on a Commit / Rollback

· ANSI closes cursors with each COMMIT causing either poor throughput (by making the server re-materialize the result set) or poor concurrency (by holding locks)

· Open Multiple Cursors on a Single Connection

· Reduces resource consumption on both client and Server

· Eliminates risk of a client-side deadlocks with itself

Special Issues for Batch Applications

ASE was not designed as a batch subsystem! It was designed as an RBDMS for large multi-user applications. Designers of batch-oriented applications should consider the following design alternatives to maximize performance :

Design Alternatives :

· Minimize Client/Server Interaction Whenever Possible

· Don't turn ASE into a "file system" by issuing single table / single row requests when, in actuality, set logic applies.

· Maximize TDS packet size for efficient Interprocess Communication (v10 only)

· New ASE 10.0 cursors declared and processed entirely within stored procedures and triggers offer significant performance gains in batch processing.

· Investigate Opportunities to Parallelize Processing

· Breaking up single processes into multiple, concurrently executing, connections (where possible) will outperform single streamed processes everytime.

· Make Use of TEMPDB for Intermediate Storage of Useful Data

Asynchronous Queries

Many, if not most, applications and 3rd Party tools are coded to send queries with the DB-Library call dbsqlexec( ) which is a synchronous call ! It sends a query and then waits for a response from ASE that the query has completed !

Designing your applications for asynchronous queries provides many benefits:

A "Cooperative" multi-tasking application design under Windows will allow users to run other Windows applications while your long queries are processed !

Provides design opportunities to parallize work across multiple ASE connections.

Implementation Choices:

· System 10 Client Library Applications:

· True asynchronous behaviour is built into the entire library. Through the appropriate use of call-backs, asynchronous behavior is the normal processing paradigm.

· Windows DB-Library Applications (not true async but polling for data):

· Use dbsqlsend(), dbsqlok(), and dbdataready() in conjunction with some additional code in WinMain() to pass control to a background process. Code samples which outline two different Windows programming approaches (a PeekMessage loop and a Windows Timer approach) are available in the Microsoft Software Library on Compuserve (GO MSL). Look for SQLBKGD.ZIP

· Non-PC DB-Library Applications (not true async but polling for data):

· Use dbsqlsend(), dbsqlok(), and dbpoll() to utilize non-blocking functions.

Generating Sequential Numbers Many applications use unique sequentially increasing numbers, often as primary keys. While there are good benefits to this approach, generating these keys can be a serious contention point if not careful. For a complete discussion of the alternatives, download Malcolm Colton's White Paper on Sequential Keys from the SQL Server Library of our OpenLine forum on Compuserve.

The two best alternatives are outlined below.

"Primary Key" Table Storing Last Key Assigned

· Minimize contention by either using a seperate "PK" table for each user table or padding out each row to a page. Make sure updates are "in-place".

· Don't include the "PK" table's update in the same transaction as the INSERT. It will serialize the transactions.

BEGIN TRAN

UPDATE pk_table SET nextkey = nextkey + 1

[WHERE table_name = @tbl_name]

COMMIT TRAN

/* Now retrieve the information */

SELECT nextkey FROM pk_table

WHERE table_name = @tbl_name]

· "Gap-less" sequences require additional logic to store and retrieve rejected values

IDENTITY Columns (v10.0 only)

· Last key assigned for each table is stored in memory and automatically included in all INSERTs (BCP too). This should be the method of choice for performance.

· Choose a large enough numeric or else all inserts will stop once the max is hit.

· Potential rollbacks in long transactions may cause gaps in the sequence !

Other Application Issues

· Transaction Logging Can Bottleneck Some High Transaction Environments

· Committing a Transaction Must Initiate a Physical Write for Recoverability

· Implementing multiple statements as a transaction can assist in these environment by minimizing the number of log writes (log is flushed to disk on commits).

· Utilizing the Client Machine's Processing Power Balances Load

· Client/Server doesn't dictate that everything be done on Server!

· Consider moving "presentation" related tasks such as string or mathematical manipulations, sorting, or, in some cases, even aggregating to the client.

· Populating of "Temporary" Tables Should Use "SELECT INTO" - balance this with dynamic creation of temporary tables in an OLTP environment. Dynamic creation may cause blocks in your tempdb.

· "SELECT INTO" operations are not logged and thus are significantly faster than there INSERT with a nested SELECT counterparts.

· Consider Porting Applications to Client Library Over Time

· True Asynchronous Behavior Throughout Library

· Array Binding for SELECTs

· Dynamic SQL

· Support for ClientLib-initiated callback functions

· Support for Server-side Cursors

· Shared Structures with Server Library (Open Server 10)

Physical Database Design Issues

· Normalized -vs- Denormalized Design

· Index Selection

· Promote "Updates-in-Place" Design

· Promote Parallel I/O Opportunities

Normalized -vs- Denormalized

· Always Start with a Completely Normalized Database

· Denormalization should be an optimization taken as a result of a performance problem

· Benefits of a normalized database include :

Accelerates searching, sorting, and index creation since tables are narrower

Allows more clustered indexes and hence more flexibility in tuning queries, since there are more tables ;

Accelerates index searching since indexes tend to be narrower and perhaps shorter ;

Allows better use of segments to control physical placement of tables ;

Fewer indexes per table, helping UPDATE, INSERT, and DELETE performance ;

Fewer NULLs and less redundant data, increasing compactness of the database ;

Accelerates trigger execution by minimizing the extra integrity work of maintaining redundant data.

Joins are Generally Very Fast Provided Proper Indexes are Available

Normal caching and cindextrips parameter (discussed in Server section) means each join will do on average only 1-2 physical I/Os.

Cost of a logical I/O (get page from cache) only 1-2 milliseconds.

There Are Some Good Reasons to Denormalize

All queries require access to the "full" set of joined data.

Majority of applications scan entire tables doing joins.

Computational complexity of derived columns require storage for SELECTs

Others ...

Index Selection

· Without a clustered index, all INSERTs and "out-of-place" UPDATEs go to the last page. The lock contention in high transaction environments would be prohibitive. This is also true for INSERTs to a clustered index on a monotonically increasing key.

· High INSERT environments should always cluster on a key which provides the most "randomness" (to minimize lock / device contention) that is usable in many queries. Note this is generally not your primary key !

· Prime candidates for clustered index (in addition to the above) include :

· Columns Accessed by a Range

· Columns Used with Order By, Group By, or Joins

· Indexes Help SELECTs and Hurt INSERTs

· Too many indexes can significantly hurt performance of INSERTs and "out-of-place" UPDATEs.

· Prime candidates for nonclustered indexes include :

· Columns Used in Queries Requiring Index Coverage

· Columns Used to Access Less than 20% (rule of thumb) of the Data.

· Unique indexes should be defined as UNIQUE to help the optimizer

· Minimize index page splits with Fillfactor (helps concurrency and minimizes deadlocks)

· Keep the Size of the Key as Small as Possible

· Accelerates index scans and tree traversals

· Use small datatypes whenever possible . Numerics should also be used whenever possible as they compare faster than strings.

Promote "Update-in-Place" Design

· "Update-in-Place" Faster by Orders of Magnitude

· Performance gain dependent on number of indexes. Recent benchmark (160 byte rows, 1 clustered index and 2 nonclustered) showed 800% difference!

· Alternative ("Out-of-Place" Update) implemented as a physical DELETE followed by a physical INSERT. These tactics result in:

Increased Lock Contention

Increased Chance of Deadlock

Decreased Response Time and Throughput

· Currently (System 10 and below), Rules for "Update-in-Place" Behavior Include :

Columns updated can not be variable length or allow nulls

Columns updated can not be part of an index used to locate the row to update

No update trigger on table being updated (because the inserted and deleted tables used in triggers get their data from the log)

In v4.9.x and below, only one row may be affected and the optimizer must know this in advance by choosing a UNIQUE index. System 10 eliminated this limitation.

Promote Parallel I/O Opportunities

· For I/O-bound Multi-User Systems, Use A lot of Logical and Physical Devices

· Plan balanced separation of objects across logical and physical devices.

· Increased number of physical devices (including controllers) ensures physical bandwidth

· Increased number of logical Sybase devices ensures minimal contention for internal resources. Look at SQL Monitor's Device I/O Hit Rate for clues. Also watch out for the 128 device limit per database.

· Create Database (in v10) starts parallel I/O on up to 6 devices at a time concurrently. If taken advantage of, expect an 800% performance gain. A 2Gb TPC-B database that took 4.5 hours under 4.9.1 to create now takes 26 minutes if created on 6 independent devices !

· Use Sybase Segments to Ensure Control of Placement

This is the only way to guarantee logical seperation of objects on devices to reduce contention for internal resources.

· Dedicate a seperate physical device and controller to the transaction log in tempdb too.

· optimize TEMPDB Also if Heavily Accessed

· increased number of logical Sybase devices ensures minimal contention for internal resources.

· systems requiring increased log throughput today must partition database into separate databases

Breaking up one logical database into multiple smaller databases increases the number number of transaction logs working in parallel.

Networking Issues

· Choice of Transport Stacks

· Variable Sized TDS Packets

· TCP/IP Packet Batching

Choice of Transport Stacks for PCs

· Choose a Stack that Supports "Attention Signals" (aka. "Out of Band Data")

· Provides for the most efficient mechanism to cancel queries.

· Essential for sites providing ad-hoc query access to large databases.

· Without "Attention Signal" capabilities (or the urgent flag in the connection string), the DB-Library functions DBCANQUERY ( ) and DBCANCEL ( ) will cause ASE to send all rows back to the Client DB-Library as quickly as possible so as to complete the query. This can be very expensive if the result set is large and, from the user's perspective, causes the application to appear as though it has hung.

· With "Attention Signal" capabilities, Net-Library is able to send an out-of-sequence packet requesting the ASE to physically throw away any remaining results providing for instantaneous response.

· Currently, the following network vendors and associated protocols support the an "Attention Signal" capable implementation:

NetManage NEWT

FTP TCP

Named Pipes (10860) - Do not use urgent parameter with this Netlib

Novell LAN Workplace v4.1 0 Patch required from Novell

Novell SPX - Implemented internally through an "In-Band" packet

Wollongong Pathway

Microsoft TCP - Patch required from Microsoft

Variable-sized TDS Packets

Pre-v4.6 TDS Does Not Optimize Network Performance Current ASE TDS packet size limited to 512 bytes while network frame sizes are significantly larger (1508 bytes on Ethernet and 4120 bytes on Token Ring).

The specific protocol may have other limitations!

For example:

· IPX is limited to 576 bytes in a routed network.

· SPX requires acknowledgement of every packet before it will send another. A recent benchmark measured a 300% performance hit over TCP in "large" data transfers (small transfers showed no difference).

· Open Client Apps can "Request" a Larger Packet Shown to have significant performance improvement on "large" data transfers such as BCP, Text / Image Handling, and Large Result Sets.

· clients:

· isql -Usa -Annnnn

· bcp -Usa -Annnnn

· ct_con_props (connection, CS_SET, CS_PACKETSIZE, &packetsize, sizeof(packetsize), NULL)

· An "SA" must Configure each Servers' Defaults Properly

· sp_configure "default packet size", nnnnn - Sets default packet size per client connection (defaults to 512)

· sp_configure "maximum packet size", nnnnn - Sets maximum TDS packet size per client connection (defaults to 512)

· sp_configure "additional netmem", nnnnn - Additional memory for large packets taken from separate pool. This memory does not come from the sp_configure memory setting.

Optimal value = ((# connections using large packets large packetsize * 3) + an additional 1-2% of the above calculation for overhead)

Each connection using large packets has 3 network buffers: one to read; one to write; and one overflow.

· Default network memory - Default-sized packets come from this memory pool.

· Additional Network memory - Big packets come this memory pool.

If not enough memory is available in this pool, the server will give a smaller packet size, down to the default

TCP/IP Packet Batching

· TCP Networking Layer Defaults to "Packet Batching"

· This means that TCP/IP will batch small logical packets into one larger physical packet by briefly delaying packets in an effort to fill the physical network frames (Ethernet, Token-Ring) with as much data as possible.

· Designed to improve performance in terminal emulation environments where there are mostly only keystrokes being sent across the network.

· Some Environments Benefit from Disabling Packet Batching

· Applies mainly to socket-based networks (BSD) although we have seen some TLI networks such as NCR's benefit.

· Applications sending very small result sets or statuses from sprocs will usually benefit. Benchmark with your own application to be sure.

· This makes ASE open all connections with the TCP_NODELAY option. Packets will be sent regardless of size.

· To disable packet batching, in pre-Sys 11, start ASE with the 1610 Trace Flag.

$SYBASE/dataserver -T1610 -d /usr/u/sybase/master.dat ...

Your errorlog will indicate the use of this option with the message:

ASE booted with TCP_NODELAY enabled.

Operating System Issues

· Never Let ASE Page Fault

· It is better to configure ASE with less memory and do more physical database I/O than to page fault. OS page faults are synchronous and stop the entire dataserver engine until the page fault completes. Since database I/O's are asynchronous, other user tasks can continue!

· Use Process Affinitying in SMP Environments, if Supported

· Affinitying dataserver engines to specific CPUs minimizes overhead associated with moving process information (registers, etc) between CPUs. Most implementations will preference other tasks onto other CPUs as well allowing even more CPU time for dataserver engines.

· Watch out for OS's which are not fully symmetric. Affinitying dataserver engines onto CPUs that are heavily used by the OS can seriously degrade performance. Benchmark with your application to find optimal binding.

· Increase priority of dataserver engines, if supported

· Give ASE the opportunity to do more work. If ASE has nothing to do, it will voluntarily yield the CPU.

· Watch out for OS's which externalize their async drivers. They need to run too!

· Use of OS Monitors to Verify Resource Usage

· The OS CPU monitors only "know" that an instruction is being executed. With ASE's own threading and scheduling, it can routinely be 90% idle when the OS thinks its 90% busy. SQL Monitor shows real CPU usage.

· Look into high disk I/O wait time or I/O queue lengths. These indicate physical saturation points in the I/O subsystem or poor data distribution.

· Disk Utilization above 50% may be subject to queuing effects which often manifest themselves as uneven response times.

· Look into high system call counts which may be symptomatic of problems.

· Look into high context switch counts which may also be symptomatic of problems.

· Optimize your kernel for ASE (minimal OS file buffering, adequate network buffers, appropriate KEEPALIVE values, etc).

· Use OS Monitors and SQL Monitor to Determine Bottlenecks

· Most likely "Non-Application" contention points include:

Resource Where to Look

--------- --------------

CPU Performance SQL Monitor - CPU and Trends

Physical I/O Subsystem OS Monitoring tools - iostat, sar...

Transaction Log SQL Monitor - Device I/O and

Device Hit Rate

on Log Device

ASE Network Polling SQL Monitor - Network and Benchmark

Baselines

Memory SQL Monitor - Data and Cache

Utilization

· Use of Vendor-support Striping such as LVM and RAID

· These technologies provide a very simple and effective mechanism of load balancing I/O across physical devices and channels.

· Use them provided they support asynchronous I/O and reliable writes.

· These approaches do not eliminate the need for Sybase segments to ensure minimal contention for internal resources.

· Non-read-only environments should expect performance degradations when using RAID levels other than level 0. These levels all include fault tolerance where each write requires additional reads to calculate a "parity" as well as the extra write of the parity data.

Hardware Configuration Issues

· Number of CPUs

· Use information from SQL Monitor to assess ASE's CPU usage.

· In SMP environments, dedicate at least one CPU for the OS.

· Advantages and scaling of VSA is application-dependent. VSA was architected with large multi-user systems in mind.

· I/O Subsystem Configuration

· Look into high Disk I/O Wait Times or I/O Queue Lengths. These may indicate physical I/O saturation points or poor data distribution.

· Disk Utilization above 50% may be subject to queuing effects which often manifest themselves as uneven response times.

· Logical Volume configurations can impact performance of operations such as create database, create index, and bcp. To optimize for these operations, create Logical Volumes such that they start on different channels / disks to ensure I/O is spread across channels.

· Discuss device and controller throughput with hardware vendors to ensure channel throughput high enough to drive all devices at maximum rating.

General ASE Tuning

· Changing Values with sp_configure or buildmaster

It is imperative that you only use sp_configure to change those parameters that it currently maintains because the process of reconfiguring actually recalculates a number of other buildmaster parameters. Using the Buildmaster utility to change a parameter "managed" by sp_configure may result in a mis-configured server and cause adverse performance or even worse ...

· Sizing Procedure Cache

· ASE maintains an MRU-LRU chain of stored procedure query plans. As users execute sprocs, ASE looks in cache for a query plan to use. However, stored procedure query plans are currently not re-entrant! If a query plan is available, it is placed on the MRU and execution begins. If no plan is in memory, or if all copies are in use, a new copy is read from the sysprocedures table. It is then optimized and put on the MRU for execution.

· Use dbcc memusage to evaluate the size and number of each sproc currently in cache. Use SQL Monitor's cache statistics to get your average cache hit ratio. Ideally during production, one would hope to see a high hit ratio to minimize the procedure reads from disk. Use this information in conjuction with your desired hit ratio to calculate the amount of memory needed.

· Memory

· Tuning memory is more a price/performance issue than anything else ! The more memory you have available, the greater than probability of minimizing physical I/O. This is an important goal though. Not only does physical I/O take significantly longer, but threads doing physical I/O must go through the scheduler once the I/O completes. This means that work on behalf of the thread may not actually continue to execute for quite a while !

· There are no longer (as of v4.8) any inherent limitations in ASE which cause a point of diminishing returns on memory size.

· Calculate Memory based on the following algorithm :

Total Memory = Dataserver Executable Size (in bytes) +
Static Overhead of 1 Mb +
User Connections x 40,960 bytes +
Open Databases x 644 bytes +
Locks x 32 bytes +
Devices x 45,056 bytes +
Procedure Cache +
Data Cache

· Recovery Interval

· As users change data in ASE, only the transaction log is written to disk right away for recoverability. "Dirty" data and index pages are kept in cache and written to disk at a later time. This provides two major benefits:

Many transactions may change a page yet only one physical write is done

ASE can schedule the physical writes "when appropriate"

· ASE must eventually write these "dirty" pages to disk.

· A checkpoint process wakes up periodically and "walks" the cache chain looking for dirty pages to write to disk

· The recovery interval controls how often checkpoint writes dirty pages.

· Tuning Recovery Interval

· A low value may cause unnecessary physical I/O lowering throughput of the system. Automatic recovery is generally much faster during boot-up.

· A high value minimizes unnecessary physical I/O and helps throughput of the system. Automatic recovery may take substantial time during boot-up.

Audit Performance Tuning for v10.0

· Potentially as Write Intensive as Logging

· Isolate Audit I/O from other components.

· Since auditing nearly always involves sequential writes, RAID Level 0 disk striping or other byte-level striping technology should provide the best performance (theoretically).

· Size Audit Queue Carefully

· Audit records generated by clients are stored in an in memory audit queue until they can be processed.

· Tune the queue's size with sp_configure "audit queue size", nnnn (in rows).

· Sizing this queue too small will seriously impact performance since all user processes who generate audit activity will sleep if the queue fills up.

· Size Audit Database Carefully

· Each audit row could require up to 416 bytes depending on what is audited.

· Sizing this database too small will seriously impact performance since all user processes who generate audit activity will sleep if the database fills up.

1.5.2: Temp Tables and OLTP

(Note from Ed: It appears that with ASE 12, Sybase have solved the problem of select/into locking the system tables for the duration of the operation. The operation is now split into two parts, the creation of the table followed byt the insert. The system tables are only locked for the first part, and so, to all intents and purposes, the operation acts like a create/insert pair whilst remaining minimally logged.

Our shop would like to inform folks of a potential problem when using temporary tables in an OLTP environment. Using temporary tables dynamically in a OLTP production environment may result in blocking (single-threading) as the number of transactions using the temporary tables increases.

Does it affect my application?

This warning only applies for SQL that is being invoked frequently in an OLTP production environment, where the use of "select into..." or "create table #temp" is common. Application using temp tables may experience blocking problems as the number of transactions increases.

This warning does not apply to SQL that may be in a report or that is not used frequently. Frequently is defined as several times per second.

Why? Why? Why?

Our shop was working with an application owner to chase down a problem they were having during peak periods. The problem they were having was severe blocking in tempdb.

What was witnessed by the DBA group was that as the number of transactions increased on this particular application, the number of blocks in tempdb also increased.

We ran some independent tests to simulate a heavily loaded server and discovered that the data pages in contention were in tempdb's syscolumns table.

This actually makes sense because during table creation entries are added to this table, regardless if it's a temporary or permanent table.

We ran another simulation where we created the tables before the stored procedure used it and the blocks went away. We then performed an additional test to determine what impact creating temporary tables dynamically would have on the server and discovered that there is a 33% performance gain by creating the tables once rather than re-creating them.

Your mileage may vary.

How do I fix this?

To make things better, do the 90's thing -- reduce and reuse your temp tables. During one application connection/session, aim to create the temp tables only once.

Let's look at the lifespan of a temp table. If temp tables are created in a batch within a connection, then all future batches and stored procs will have access to such temp tables until they're dropped; this is the reduce and reuse strategy we recommend. However, if temp tables are created in a stored proc, then the database will drop the temp tables when the stored proc ends, and this means repeated and multiple temp table creations; you want to avoid this.

Recode your stored procedures so that they assume that the temporary tables already exist, and then alter your application so that it creates the temporary tables at start-up -- once and not every time the stored procedure is invoked.

That's it! Pretty simple eh?

Summary

The upshot is that you can realize roughly a 33% performance gain and not experience the blocking which is difficult to quantify due to the specificity of each application.

Basically, you cannot lose.

Solution in pseudo-code

If you have an application that creates the same temp table many times within one connection, here's how to convert it to reduce and reuse temp table creations. Raymond Lew has supplied a detailed example for trying this.

Old

open connection

loop until time to go

exec procedure vavoom_often

/* vavoom_often creates and uses #gocart for every call */

/* eg: select * into #gocart from gocart */

loop-end

close connection

New

open connection

/* Create the temporary table outside of the sproc */

select * into #gocart from gocart where 1 =2 ;

loop until time to go

exec procedure vavoom_often

/* vavoom_often reuses #gocart which */

/* was created before exec of vavoom_often */

/* - First statement may be a truncate table #gocart */

/* - Execute with recompile */

/* if your table will have more than 10 data pages */

/* as the optimizer will assume 10 data pages for temp tables */

loop-end

close connection

Note that it is necessary to call out the code to create the table and it becomes a pain in the butt because the create-table statement will have to be replicated in any stored proc and in the initialization part of the application - this can be a maintenance nuisance. This can be solved by using any macro package such as m4 or cpp. or by using and adapting the scripts from Raymond Lew.

Brian Black posted a stronger notice than this to the SYBASE-L list, and I would agree, that any use of select/into in a production environments should looked at very hard. Even in DSS environments, especially if they share tempdb with an OLTP environment, should use select/into with care.

From: Raymond Lew

At our company, we try to keep the database and the application loosely coupled to allow independent changes at the frontend or the backend as long as the interface stays the same. Embedding temp table definitions in the frontend would make this more difficult.

To get away from having to embed the temp table definitions in the frontend code, we are storing the temp table definitions in the database. The frontend programs retrieve the definitions and declare the tables dynamically at the beginning of each session. This allows for the change of backend procedures without changes in the frontend when the API does not change.

Enclosed below are three scripts. The first is an isql script to create the tables to hold the definitions. The second is a shell script to set up a sample procedure named vavoom. The third is shell script to demonstrate the structure of application code.

I would like to thank Charles Forget and Gordon Rees for their assistance on these scripts.

--start of setup------------------------------------------------------

/* Raymond Lew - 1996-02-20 */

/* This isql script will set up the following tables:

gocart - sample table

app_temp_defn - where temp table definitions are stored

app_temp_defn_group - a logical grouping of temp table definitions

for an application function

/******************************/

/* gocart table - sample table*/

/******************************/

drop table gocart

create table gocart

(

cartname char(10) null

,cartcolor char(30) null

)

create unique clustered index gocart1 on gocart (cartname)

insert into gocart values ('go1','blue ')

insert into gocart values ('go2','pink ')

insert into gocart values ('go3','green ')

insert into gocart values ('go4','red ')

/****************************************************************/

/* app_temp_defn - definition of temp tables with their indexes */

/****************************************************************/

drop table app_temp_defn

create table app_temp_defn

(

/* note: temp tables are unique only in first 13 chars */

objectname char(20) not null

,seq_no smallint not null

,defntext char(255) not null

)

create unique clustered index app_temp_defn1

on app_temp_defn (objectname,seq_no)

insert into app_temp_defn

values ('#gocart',1,'select * into #gocart')

insert into app_temp_defn

values ('#gocart',2,' from gocart where 1=2 ')

insert into app_temp_defn

values ('#gocartindex',1,

"create unique index gocartindex on #gocart (cartname) ")

insert into app_temp_defn

values ('#gocart1',1, 'select * into #gocart1 from gocart where 1=2')

/***********************************************************************/

/* app_temp_defn_group - groupings of temp definitions by applications */

/***********************************************************************/

drop table app_temp_defn_group

create table app_temp_defn_group

(

appname char(8) not null

,objectname char(20) not null

)

create unique clustered index app_temp_defn_group1

on app_temp_defn_group (appname,objectname)

insert into app_temp_defn_group values('abc','#gocart')

insert into app_temp_defn_group values('abc','#gocartindex')

/***********************************************************/

/* get_temp_defn - proc for getting the temp defn by group */

/***********************************************************/

drop procedure get_temp_defn

create procedure get_temp_defn

(

@appname char(8)

)

if @appname = ''

select defntext

from app_temp_defn

order by objectname, seq_no

else

select defntext

from app_temp_defn a

, app_temp_defn_group b

where a.objectname = b.objectname

and b.appname = @appname

order by a.objectname, a.seq_no

return

/* let's try some tests */

exec get_temp_defn ''

exec get_temp_defn 'abc'

--end of setup --------------------------------------------------

--- start of make.vavoom --------------------------------------------

#!/bin/sh

# Raymond Lew - 1996-02-20

# bourne shell script for creating stored procedures using

# app_temp_defn table

# demo procedure vavoom created here

# note: you have to change the passwords, id and etc. for your site

# note: you might have to some inline changes to make this work

# check out the notes within the body

# get the table defn's into a text file

# note: next line :you will need to end the line immediately after eot \

isql -Ukryten -Pjollyguy -Sstarbug -w255 << eot \

| grep -v '\-\-\-\-' | grep -v 'defntext ' | grep -v ' affected' > tabletext

exec get_temp_defn ''

eot

# note: prev line :you will need to have a newline immediately after eot

# go mess around in vi

vi tabletext

# create the proc vavoom after running the temp defn's into db

isql -Ukryten -Pjollyguy -Sstarbug -e << eot |more

`cat tabletext`

drop procedure vavoom

create procedure vavoom

(

@color char(10)

)

truncate table #gocart1 /* who knows what lurks in temp tables */

if @color = ''

insert #gocart1 select * from gocart

else

insert #gocart1 select * from gocart where cartcolor=@color

select @color '@color', * from #gocart1

return

exec vavoom ''

exec vavoom 'blue'

eot

# note: prev line :you will need to have a newline immediately after eot

exit

# end of unix script

--- end of make.vavoom --------------------------------------------

--- start of defntest.sh -------------------------------------------

#!/bin/sh

# Raymond Lew 1996-02-01

# test script: demonstrate with a bourne shell how an application

# would use the temp table definitions stored in the database

# note: you must run setup and make.vavoom first

# note: you have to change the passwords, id and etc. for your site

# note: you might have to some inline changes to make this work

# check out the notes within the body

# get the table defn's into a text file

# note: next line :you will need to end the line immediately after eot \

isql -Ukryten -Pjollyguy -Sstarbug -w255 << eot \

| grep -v '\-\-\-\-' | grep -v 'defntext ' | grep -v ' affected' > tabletext

exec get_temp_defn ''

eot

# note: prev line :you will need to have a newline immediately after eot

# go mess around in vi

vi tabletext

isql -Ukryten -Pjollyguy -Sstarbug -e << eot | more

`cat tabletext`

exec vavoom ''

exec vavoom 'blue'

eot

# note: prev line :you will need to have a newline immediately after eot

exit

# end of unix script

--- end of defntest.sh -------------------------------------------

That's all, folks. Have Fun

1.5.3: Differences between clustered and non-clustered

Preface

I'd like to talk about the difference between a clustered and a non-clustered index. The two are very different and it's very important to understand the difference between the two to in order to know when and how to use each.

I've pondered hard to find the best analogy that I could think of and I've come up with ... the phone book. Yes, a phone book.

Imagine that each page in our phone book is equivalent to a Sybase 2K data page. Every time we read a page from our phone book it is equivalent to one disk I/O.

Since we are imagining, let's also imagine that our mythical ASE (that runs against the phone book) has only enough data cache to buffer 200 phone pages. When our data cache gets full we have to flush an old page out so we can read in a new one.

Fasten your seat belts, because here we go...

Clustered Index

A phone book lists everyone by last name. We have an A section, we have a B section and so forth. Within each section my phone book is clever enough to list the starting and ending names for the given page.

The phone book is clustered by last name.

create clustered index on phone_book (last_name)

It's fast to perform the following queries on the phone book:

· Find the address of those whose last name is Cisar.

· Find the address of those whose last name is between Even and Fa

Searches that don't work well:

· Find the address of those whose phone number is 440-1300.

· Find the address of those whose prefix is 440

In order to determine the answer to the two above we'd have to search the entire phone book. We can call that a table scan.

Non-Clustered Index

To help us solve the problem above we can build a non-clustered index.

create nonclustered index on phone_book (phone_number)

Our non-clustered index will be built and maintained by our Mythical ASE as follows:

Create a data structure that will house a phone_number and information where the phone_number exists in the phone book: page number and the row within the page.

The phone numbers will be kept in ascending order.

Scan the entire phone book and add an entry to our data structure above for each phone number found.

For each phone number found, note along side it the page number that it was located and which row it was in.

any time we insert, update or delete new numbers, our M-ASE will maintain this secondary data structure. It's such a nice Server.

Now when we ask the question:

Find the address of those whose phone number is 440-1300

we don't look at the phone book directly but go to our new data structure and it tells us which page and row within the page the above phone number can be found. Neat eh?

Draw backs? Well, yes. Because we probably still can't answer the question:

Find the address of those whose prefix is 440

This is because of the data structure being used to implement non-clustered indexes. The structure is a list of ordered values (phone numbers) which point to the actual data in the phone book. This indirectness can lead to trouble when a range or a match query is issued.

The structure may look like this:

------------------------------------

|Phone Number | Page Number/Row |

====================================

| 440-0000 | 300/23 |

| 440-0001 | 973/45 |

| 440-0002 | 23/2 |

| ... | |

| 440-0030 | 973/45 |

| 440-0031 | 553/23 |

| ... | |

------------------------------------

As one can see, certain phone numbers may map to the same page. This makes sense, but we need to consider one of our constraints: our Server only has room for 200 phone pages.

What may happen is that we re-read the same phone page many times. This isn't a problem if the phone page is in memory. We have limited memory, however, and we may have to flush our memory to make room for other phone pages. So the re-reading may actually be a disk I/O.

The Server needs to decide when it's best to do a table scan versus using the non-clustered index to satisfy mini-range type of queries. The way it decides this is by applying a heuristic based on the information maintained when an update statistics is performed.

In summary, non-clustered indexes work really well when used for highly selective queries and they may work for short, range type of queries.

Suggested Uses

Having suffered many table corruption situations (with 150 ASEs who wouldn't? :-)), I'd say always have a clustered index. With a clustered index you can fish data out around the bad spots on the table thus having minimal data loss.

When you cluster, build the cluster to satisfy the largest percentage of range type queries. Don't put the clustered index on your primary key because typically primary keys are increasing linearly. What happens is that you end up inserting all new rows at the end of the table thus creating a hot spot on the last data page.

For detail rows, create the clustered index on the commonly accessed foreign key. This will aid joins from the master to it.

Use nonclustered index to aid queries where your selection is very selective. For example, primary keys. :-)

1.5.4: Optimistic versus Pessimistic locking?

This is the same problem another poster had ... basically locking a record to ensure that it hasn't changed underneath ya.

fcasas@ix.netcom.com has a pretty nifty solution if you are using ct-lib (I'll include that below -- hope it's okay Francisco ... :-)) ...

Basically the problem you are facing is one of being a pessimist or an optimist.

I contend that your business really needs to drive this.

Most businesses (from my experience) can be optimistic.

That is, if you are optimistic that the chances that someone is going to change something from underneath the end-user is low, then do nothing about it.

On the other hand, if you are pessimistic that someone may change something underneath the end-user, you can solve it at least as follows:

Solution #1

Use a timestamp on a header table that would be shared by the common data. This timestamp field is a Sybase datatype and has nothing to do with the current time. Do not attempt to do any operations on this column other than comparisons. What you do is when you grab data to present to the end-user, have the client software also grab the timestamp column value. After some thing time, if the end-user wishes to update the database, compare the client timestamp with what's in the database and it it's changed, then you can take appropriate action: again this is dictated by the business.

Problem #1

If users are sharing tables but columns are not shared, there's no way to detect this using timestamps because it's not sufficiently granular.

Solution #2 (presented by fcasas)

... Also are you coding to ct-lib directly? If so there's something that you could have done, or may still be able to do if you are using cursors.

With ct-lib there's a ct_describe function that lets you see key data. This allows you to implement optimistic locking with cursors and not need timestamps. Timestamps are nice, but they are changed when any column on a row changes, while the ct_describe mechanism detects changes at the columns level for a greater degree of granularity of the change. In other words, the timestamp granularity is at the row, while ct_describes CS_VERSION_KEY provides you with granularity at the column level.

Unfortunately this is not well documented and you will have to look at the training guide and the manuals very closely.

Further if you are using cursors do not make use of the

[for {read only | update [of column_name_list]}]

of the select statement. Omitting this clause will still get you data that can still be updated and still only place a shared lock on the page. If you use the read only clause you are acquiring shared locks, but the cursor is not updatable. However, if you say

update [of ...

will place updated locks on the page, thus causing contention. So, if you are using cursors don't use the above clause. So, could you answer the following three questions:

Are you using optimistic locking?

Are you coding to ct-lib?

Are you using cursors?

Problem #2

You need to be coding with ct-lib ...

Solution #3

Do nothing and be optimistic. We do a lot of that in our shop and it's really not that big of a problem.

Problem #3

Users may clobber each other's changes ... then they'll come looking for you to clobber you! :-)

1.5.5: How do I force an index to be used?

System 11

In System 11, the binding of the internal ordinal value is alleviated so that instead of using the ordinal index value, the index name can be used instead:

select ... from my_table (index my_first_index)

Sybase 4.x and Sybase System 10

All indexes have an ordinal value assigned to them. For example, the following query will return the ordinal value of all the indexes on my_table:

select name, indid
from sysindexes
where id = object_id("my_table")

Assuming that we wanted to force the usuage of index numbered three:

select ... from my_table(3)

Note: using a value of zero is equivalent to forcing a table scan. Whilst this sounds like a daft thing to do, sometimes a table scan is a better solution than heavy index scanning.

It is essential that all index hints be well documented. This is good DBA practice. It is especially true for Sybase System 10 and below.

One scheme that I have used that works quite well is to implement a table similar to sysdepends in the database that contains the index hints.

create table idxdepends

(

tblname varchar(32) not null -- Table being hinted

,depname varchar(50) not null -- Proc, trigger or app that

-- contains hint.

,idxname varchar(32) not null -- Index being hinted at

--,hintcount int null -- You may want to count the

-- number of hints per proc.

)

Obviously it is a manual process to keep the table populated, but it can save a lot of trouble later on.

1.5.6: Why place tempdb and log on low numbered devices?

System 10 and below.

In System 10 and Sybase 4.X, the I/O scheduler starts at logical device (ldev) zero and works up the ldev list looking for outstanding I/O's to process. Taking this into consideration, the following device fragments (disk init) should be added before any others:

tempdb

log

1.5.7: How much memory to configure?

System 10 and below.

Overview

At some point you'll wonder if your ASE has been configured with sufficient memory. We hope that it's not during some crisis but that's probably when it'll happen.

The most important thing in setting up memory for a ASE is that it has to be large enough to accommodate:

· concurrent user connections

· active procedures

· and concurrent open databases.

By not setting the ASE up correctly it will affect the performance of it. A delicate balance needs to be struck where your ASE is large enough to accommodate the users but not too large where it adversely affects the CPU Server (such as causing swapping).

Assumptions made of the reader:

· The reader has some experience administering ASEs.

· All queries have been tuned and that there are no unnecessary table scans.

Preface

As the ASE starts up, it pre-allocates its structures to support the configuration. The memory that remains after the pre-allocation phase is the available cache.

The available cache is partitioned into two pieces:

buffer cache - data pages to be sent to a user connection or flushed to disk.

procedure cache - where query plans live.

The idea is to determine if the buffer cache and the procedure cache are of adequate size. As a DBA you can use dbcc memusage to ascertain this.

The information provided from a dbcc memusage, daunting at first, but taken in sections, is easy to understand and provides the DBA with the vital information that is necessary to determine if more memory is required and where it is required.

If the procedure cache is too small, user connections will get sporadic 701's:

There is insufficient system memory to run this query.

If the buffer cache is too small, response time may be poor or spiky.

The following text describes how to interpret the output of dbcc memusage and to correlate this back to the fundamental question:

Does my ASE have enough memory?

Definitions

Before delving into the world of dbcc memusage some definitions to get us through.

Buffer Cache (also referred to as the Data Cache)

Area of memory where ASE stores the most recently used data pages and index pages in 2K page units. If ASE finds a data page or index page in the buffer cache, it doesn't need to perform a physical I/O (it is reported as a logical I/O). If a user connection selects data from a database, the ASE loads the 2K data page(s) here and then hands the information off to the user connection. If a user connection updates data, these pages are altered, and then they are flushed out to disk by the ASE.

This is a bit simplistic but it'll do. Read on for more info though.

The cache is maintained as a doubly linked list. The head of the list is where the most recently used pages are placed. Naturally towards the tail of the chain are the least recently used pages. If a page is requested and it is found on the chain, it is moved back to the front of the chain and the information is relayed, thus saving a physical I/O.

But wait! this recycling is not done forever. When a checkpoint occurs any dirty pages are flushed. Also, the parameter cbufwashsize determines how many times a page containing data can be recycled before it has to be flushed out to disk. For OAM and index pages the following parameters apply coamtrips and cindextrips respectively.

Procedure Cache

Area of memory where ASE stores the most recently used query plans of stored procedures and triggers. This procedure cache is also used by the Server when a procedure is being created and when a query is being compiled. Just like the buffer cache, if SQL Server finds a procedure or a compilation already in this cache, it doesn't need to read it from the disk.

The size of procedure cache is determined by the percentage of remaining memory configured for this Server parameter after ASE memory needs are met.

Available Cache

When the ASE starts up it pre-allocates its data structures to support the current configuration. For example, based on the number of user connections, additional netmem, open databases and so forth the dataserver pre-allocates how much memory it requires to support these configured items.

What remains after the pre-allocation is the available cache. The available cache is divided into buffer cache and procedure cache. The sp_configure "procedure cache" parameter determines the percentage breakdown. A value of 20 would read as follows:

20% of the available cache is dedicated to the procedure cache and 80% is dedicated to the buffer cache.

Your pal: dbcc memusage

dbcc memusage takes a snapshot of your ASE's current memory usage and reports this vital information back to you. The information returned provides information regarding the use of your procedure cache and how much of the buffer cache you are currently using.

An important piece of information is the size of the largest query plan. We'll talk about that more below.

It is best to run dbcc memusage after your ASE has reached a working set. For example, at the end of the day or during lunch time.

Running dbcc memusage will freeze the dataserver while it does its work. The more memory you have configured for the ASE the longer it'll take. Our experience is that for a ASE with 300MB it'll take about four minutes to execute. During this time, nothing else will execute: no user queries, no sp_who's...

In order to run dbcc memusage you must have sa privileges. Here's a sample execution for discussion purposes:

1> /* send the output to the screen instead of errorlog */

2> dbcc traceon(3604)

3> go

1> dbcc memusage

2> go

Memory Usage:

Meg. 2K Blks Bytes

Configured Memory:300.0000 153600 314572800

Code size: 2.6375 1351 2765600

Kernel Structures: 77.6262 39745 81396975

Server Structures: 54.4032 27855 57045920

Page Cache:129.5992 66355 135894640

Proc Buffers: 1.1571 593 1213340

Proc Headers: 25.0840 12843 26302464

Number of page buffers: 63856

Number of proc buffers: 15964

Buffer Cache, Top 20:

DB Id Object Id Index Id 2K Buffers

6 927446498 0 9424

6 507969006 0 7799

6 959446612 0 7563

6 116351649 0 7428

6 2135014687 5 2972

6 607445358 0 2780

6 507969006 2 2334

6 2135014687 0 2047

6 506589013 0 1766

6 1022066847 0 1160

6 116351649 255 987

6 927446498 8 897

6 927446498 10 733

6 959446612 7 722

6 506589013 1 687

6 971918604 0 686

6 116351649 6 387

Procedure Cache, Top 20:

Database Id: 6

Object Id: 1652357121

Object Name: lp_cm_case_list

Version: 1

Uid: 1

Type: stored procedure

Number of trees: 0

Size of trees: 0.000000 Mb, 0.000000 bytes, 0 pages

Number of plans: 16

Size of plans: 0.323364 Mb, 339072.000000 bytes, 176 pages

----

Database Id: 6

Object Id: 1668357178

Object Name: lp_cm_subcase_list

Version: 1

Uid: 1

Type: stored procedure

Number of trees: 0

Size of trees: 0.000000 Mb, 0.000000 bytes, 0 pages

Number of plans: 10

Size of plans: 0.202827 Mb, 212680.000000 bytes, 110 pages

----

Database Id: 6

Object Id: 132351706

Object Name: csp_get_case

Version: 1

Uid: 1

Type: stored procedure

Number of trees: 0

Size of trees: 0.000000 Mb, 0.000000 bytes, 0 pages

Number of plans: 9

Size of plans: 0.149792 Mb, 157068.000000 bytes, 81 pages

----

Database Id: 6

Object Id: 1858261845

Object Name: lp_get_last_caller_new

Version: 1

Uid: 1

Type: stored procedure

Number of trees: 0

Size of trees: 0.000000 Mb, 0.000000 bytes, 0 pages

Number of plans: 2

Size of plans: 0.054710 Mb, 57368.000000 bytes, 30 pages

...

1> /* redirect output back to the errorlog */

2> dbcc traceoff(3604)

3> go

Dissecting memusage output

The output may appear overwhelming but it's actually pretty easy to parse. Let's look at each section.

Memory Usage

This section provides a breakdown of the memory configured for the ASE.

Memory Usage:

Meg. 2K Blks Bytes

Configured Memory:300.0000 153600 314572800

Code size: 2.6375 1351 2765600

Kernel Structures: 77.6262 39745 81396975

Server Structures: 54.4032 27855 57045920

Page Cache:129.5992 66355 135894640

Proc Buffers: 1.1571 593 1213340

Proc Headers: 25.0840 12843 26302464

Number of page buffers: 63856

Number of proc buffers: 15964

The Configured Memory does not equal the sum of the individual components. It does in the sybooks example but in practice it doesn't always. This is not critical and it is simply being noted here.

The Kernel Structures and Server structures are of mild interest. They can be used to cross-check that the pre-allocation is what you believe it to be. The salient line items are Number of page buffers and Number of proc buffers.

The Number of proc buffers translates directly to the number of 2K pages available for the procedure cache.

The Number of page buffers is the number of 2K pages available for the buffer cache.

As a side note and not trying to muddle things, these last two pieces of information can also be obtained from the errorlog:

... Number of buffers in buffer cache: 63856.

... Number of proc buffers allocated: 15964.

In our example, we have 15,964 2K pages (~32MB) for the procedure cache and 63,856 2K pages (~126MB) for the buffer cache.

Buffer Cache

The buffer cache contains the data pages that the ASE will be either flushing to disk or transmitting to a user connection.

If this area is too small, the ASE must flush 2K pages sooner than might be necessary to satisfy a user connection's request.

For example, in most database applications there are small edit tables that are used frequently by the application. These tables will populate the buffer cache and normally will remain resident during the entire life of the ASE. This is good because a user connection may request validation and the ASE will find the data page(s) resident in memory. If however there is insufficient memory configured, then these small tables will be flushed out of the buffer cache in order to satisfy another query. The next time a validation is requested, the tables will have to be re-read from disk in order to satisfy the request. Your performance will degrade.

Memory access is easily an order of magnitude faster than performing a physical I/O.

In this example we know from the previous section that we have 63,856 2K pages (or buffers) available in the buffer cache. The question to answer is, "do we have sufficient buffer cache configured?"

The following is the output of the dbcc memusage regarding the buffer cache:

Buffer Cache, Top 20:

DB Id Object Id Index Id 2K Buffers

6 927446498 0 9424

6 507969006 0 7799

6 959446612 0 7563

6 116351649 0 7428

6 2135014687 5 2972

6 607445358 0 2780

6 507969006 2 2334

6 2135014687 0 2047

6 506589013 0 1766

6 1022066847 0 1160

6 116351649 255 987

6 927446498 8 897

6 927446498 10 733

6 959446612 7 722

6 506589013 1 687

6 971918604 0 686

6 116351649 6 387

Index Legend

Value	Definition
0	Table data
1	Clustered index
2-250	Nonclustered indexes
255	Text pages

· To translate the DB Id use select db_name(#) to map back to the database name.

· To translate the Object Id, use the respective database and use the select object_name(#) command.

It's obvious that the first 10 items take up the largest portion of the buffer cache. Sum these values and compare the result to the amount of buffer cache configured.

Summing the 10 items nets a result of 45,263 2K data pages. Comparing that to the number of pages configured, 63,856, we see that this ASE has sufficient memory configured.

When do I need more Buffer Cache?

I follow the following rules of thumb to determine when I need more buffer cache:

· If the sum of all the entries reported is equal to the number of pages configured and all entries are relatively the same size. Crank it up.

· Note the natural groupings that occur in the example. If the difference between any of the groups is greater than an order of magnitude I'd be suspicious. But only if the sum of the larger groups is very close to the number of pages configured.

Procedure Cache

If the procedure cache is not of sufficient size you may get sporadic 701 errors:

There is insufficient system memory to run this query.

In order to calculate the correct procedure cache one needs to apply the following formula (found in ASE Troubleshooting Guide - Chapter 2, Procedure Cache Sizing):

proc cache size = max(# of concurrent users) * (size of the largest plan) * 1.25

The flaw with the above formula is that if 10% of the users are executing the largest plan, then you'll overshoot. If you have distinct classes of connections whose largest plans are mutually exclusive then you need to account for that:

ttl proc cache = proc cache size * x% + proc cache size * y% ...

The max(# of concurrent users) is not the number of user connections configured but rather the actual number of connections during the peak period.

To compute the size of the largest [query] plan take the results from the dbcc memusage's, Procedure Cache section and apply the following formula:

query plan size = [size of plans in bytes] / [number of plans]

We can compute the size of the query plan for lp_cm_case_list by using the output of the dbcc memusage:

...

Database Id: 6

Object Id: 1652357121

Object Name: lp_cm_case_list

Version: 1

Uid: 1

Type: stored procedure

Number of trees: 0

Size of trees: 0.000000 Mb, 0.000000 bytes, 0 pages

Number of plans: 16

Size of plans: 0.323364 Mb, 339072.000000 bytes, 176 pages

----

...

Entering the respective numbers, the query plan size for lp_cm_case_list is 21K:

query plan size = 339072 / 16
query plan size = 21192 bytes or 21K

The formula would be applied to all objects found in the procedure cache and the largest value would be plugged into the procedure cache size formula:

Query Plan Sizes

Object	Query Plan Size
lp_cm_case_list	21K
lp_cm_subcase_list	21K
csp_get_case	19K
lp_get_last_caller_new	28K

The size of the largest [query] plan is 28K.

Entering these values into the formula:

proc cache size = max(# of concurrent users) * (size of the largest plan) * 1.25
proc cache size = 491 connections * 28K * 1.25
proc cache size = 17,185 2K pages required

Our example ASE has 15,964 2K pages configured but 17,185 2K pages are required. This ASE can benefit by having more procedure cache configured.

This can be done one of two ways:

If you have some headroom in your buffer cache, then sp_configure "procedure cache" to increase the ratio of procedure cache to buffer cache or

procedure cache =

[ proposed procedure cache ] /
( [ current procedure cache ] + [ current buffer cache ] )

The new procedure cache would be 22%:

procedure cache = 17,185 / ( 15,964 + 63,856 )
procedure cache = .2152 or 22%

If the buffer cache cannot be shrunken, then sp_configure "memory" to increase the total memory:

mem size =

([ proposed procedure cache ]) /
([ current procedure cache ] / [ current configured memory ])

The new memory size would be 165,399 2K pages, assuming that the procedure cache is unchanged:

mem size = 17,185 / ( 15,964 / 153,600 )
mem size = 165,399 2K pages

1.5.8: Why should I use stored procedures?

There are many advantages to using stored procedures (unfortunately they do not handle the text/image types):

· Security - you can revoke access to the base tables and only allow users to access and manipulate the data via the stored procedures.

· Performance - stored procedures are parsed and a query plan is compiled. This information is stored in the system tables and it only has to be done once.

· Network - if you have users who are on a WAN (slow connection) having stored procedures will improve throughput because less bytes need to flow down the wire from the client to ASE.

· Tuning - if you have all your SQL code housed in the database, then it's easy to tune the stored procedure without affecting the clients (unless of course the parameter change).

· Modularity - during application development, the application designer can concentrate on the front-end and the DB designer can concentrate on the ASE.

· Network latency - a client on a LAN may seem slower if it is sending large numbers of separate requests to a database server, bundling them into one procedure call may improve responsiveness. Also, servers handling large numbers of small requests can spend a surprising amount of CPU time performing network IO.

· Minimise blocks and deadlocks - it is a lot easier to handle a deadlock if the entire transaction is performed in one database request, also locks will be held for a shorter time, improving concurrency and potentially reducing the number of deadlocks. Further, it is easier to ensure that all tables are accessed in a consistent order if code is stored centrally rather than dispersed among a number of apps.

1.5.9: You and showplan output

As recently pointed out in the Sybase-L list, the showplan information that was here is terribly out of date. It was written back when the output from ASE and MS SQL Server were identical. (To see just how differenet they have become, have a look at the O'Reilly book "Transact-SQL Programming". It does a line for line comparison.) The write up in the Performance and Tuning Guide is excellent, and this section was doing nothing but causing problems.

If you do have a need for the original document, then it can be found here, but it will no longer be considered part of the official FAQ.

1.5.10: Poor man's sp_sysmon

This is needed for System 10 and Sybase 4.9.2 where there is no sp_sysmon command available.

Fine tune the waitfor for your application. You may need TS Role -- see Q3.1.

use master

dbcc traceon(3604)

dbcc monitor ("clear", "all", "on")

waitfor delay "00:01:00"

dbcc monitor ("sample", "all", "on")

dbcc monitor ("select", "all", "on")

dbcc traceon(8399)

select field_name, group_name, value

from sysmonitors

dbcc traceoff(8399)

dbcc traceoff(3604)

1.5.11: View MRU-LRU procedure cache chain

dbcc procbuf gives a listing of the current contents of the procedure cache. By repeating the process at intervals it is possible to watch procedures moving down the MRU-LRU chain, and so to see how long procedures remain in cache. The neat thing about this approach is that you can size your cache according to what is actually happening, rather than relying on estimates based on assumptions that may not hold on your site.

To run it:

dbcc traceon(3604)

dbcc procbuf

If you use sqsh it's a bit easier to grok the output:

dbcc traceon(3604);

dbcc procbuf;|fgrep <pbname>

See Q1.5.7 regarding procedure cache sizing.

1.5.12: Improving Text/Image Type Performance

If you know that you are going to be using a text/insert column immediately, insert the row setting the column to a non-null value.

There's a noticeable performance gain.

Unfortunately, text and image datatypes cannot be passed as parameters to stored procedures. The address of the text or image location must be created and returned where it is then manipulated by the calling code. This means that transactions involving both text and image fields and stored procedures are not atomic. However, the datatypes can still be declared as not null in the table definition.

Given this example -

create table key_n_text

(

key int not null,

notes text not null

)

This stored procedure can be used -

create procedure sp_insert_key_n_text

@key int,

@textptr varbinary(16) output

** Generate a valid text pointer for WRITETEXT by inserting an

** empty string in the text field.

insert key_n_text

(

key,

notes

)

values

(

@key,

)

select @textptr = textptr(notes)

from key_n_text

where key = @key

return 0

The return parameter is then used by the calling code to update the text field, via the dbwritetext() function if using DB-Library for example.

Sunday, November 21, 2010

Sybase Adaptive Server Enterprise

No comments:

Post a Comment