Table of Contents

1. Overview
2. General
3. Server

4. Client
 

1. Overview

This guide is intended as a supplement to the set of publications that is available with the release of Sybase Adaptive Server Enterprise 11.x (ASE 11.x). Its purpose is to provide some guidelines when developing and tuning applications that run on Adaptive Server Enterprise 11.x. This guide, by no means, replaces the need to attend the Performance & Tuning course given by Sybase Education. This course is a pre-requisite to any serious performance and tuning analysis. This guide is meant to compliment the Server publications and course material, and in fact, in many places great care has been taken not to be redundant with materials.

A new document in the Adaptive Server Enterprise 11.x set is the Performance and Tuning Guide (P&T Guide). It is expected that the reader of this guide has already read that document and is familiar with its concepts. This guide is intended to focus on some of the specific areas that developers of third party applications run across.

We also strongly advise you to (periodically) visit the Sybase World Wide Web pages for the latest information, whitepapers, performance tips, sample programs, etc. Specifically look at:
 

http://www.sybase.com

http://www.sybase.com/products/databases/sql_server/information/performance.html

http://www.sybase.com/products/samples/
 

Our goal is to help you achieve what you want of your application - to be fast (as measured by response time and throughput) and to be able to support more current users.

Being a developer of applications that will be deployed at your customers sites presents you with 3 challenges the typical end-user is spared (not everyone in the audience will face all 3):

the application was originally architected as a mainframe application, having been (minimally) re-written for client/server technology and for relational databases (i.e. - it uses the database to store data but doesn't apply the power and technologies offered); and

you need to integrate your application with the RDBMS as seamlessly as possible, (this especially applies to backup and system administrative functions).

We realize that many of you have not chosen Sybase as your primary platform and accordingly have written your application with another architecture in mind. This can often lead to inefficient practices on Sybase which we will try to point out and present feasible solutions.

We will concentrate on two areas, namely concurrency (locking and deadlocking issues) and performance (once we get it to run, making it run fast).

The table below is a summary of the topics/solutions that are discussed in more depth in subsequent sections. We included a rating (High - Medium - Low) for each on the benefit that might be gained from implementation and for the coding effort you might expect.
 
 
 
 

2. General

 The architecture of many of the applications written by the enterprise solutions developers for Sybase today follow an approach where the client and the Adaptive Server Enterprise reside on the same machine.

Adaptive Server Enterprise was not designed to share resources with the application server.
 
 

Ideally, whether or not your application resides on the same machine as the server, you want to minimize the traffic between the application and the Adaptive Server Enterprise . This means:· Doing the work as close to the data as possible. Let the server do the work through use of stored procedures, triggers, joins and other set oriented processing wherever possible. Avoid processing one row at a time.

· Sending multiple commands in a single batch instead of sending each command independently.

· Filtering the results set before it is returned, eliminating unnecessary columns and rows.  You also want to build the application as optimally as possible. Keep the executable as small as possible and compile with optimization flags. Link only those libraries necessary.
 
 

Of course, the place where you can gain the most performance improvement is in your query design, transaction design, and database design. If you have the flexibility to modify the design of your database, pay particular attention to chapter 2 of the P&T Guide and the discussions of collapsing and splitting tables.
 

uning for Performance

Step 1 - Knowing Your Application

 There are two parts to knowing your application. The first is understanding how the application is to be used. Which tables get hit most frequently, what is the mix of transactions, are there certain hours when the database is in an insert/update intensive mode, other periods when reporting is the major activities, are there somewhat quiet times? Knowing how the data is to be used allows you to decide whether or not some tables should be de-normalized (adding redundant columns to eliminate joins, adding derived columns so that aggregate values are readily available), and/or whether some tables should be duplicated (replicated) or split.
 
 

The other part is the empirical knowledge you gain after collecting and examining all the statistics either from a running system or, better yet, a controlled, repeatable simulation environment. Here's where data from the profiled application is captured through a variety of tools (operating system monitoring tools, Sybase SQL monitoring tools (i.e. SQL Monitor and "dbcc monitor"), third party monitoring tools, and traces and statistics collected through the application). (Because you, the developer of this database independent API don't have any control or, sometimes, knowledge of what the application is sending, if you can build into the application tools to capture important information, like the flow and frequency of statements and how long they take to process, you'll be way ahead of the game in being able to tune your application.)  Step 2 - Knowing What's Tunable

 - packet size, frame size and other networking parameters

- server configuration parameters, i.e. cschedspins, memory, cache

- application configurations

- number of cursor rows fetched at a time  Step 3 - Having the Right Testing Environment

 This includes being able to re-run the benchmarks at will (after first establishing the baseline). It means having a benchmark that approximates reality (containing a transaction mix, multiple users, with appropriate timing, that simulates what the real world will offer). And it means having the results of the tests being easily understood and compared (this means you can't stand there with a stopwatch - you must build the timings into the application).
 
 

There are testing tools that are available that can help you build a multi-user testing environment. The following in not a conclusive list by any means (or an endorsement) but a starting place of vendors who have offering: Performix, Mercury, and Neal Nelson.  Step 4 - What to Monitor

 At the OS level: (use tools such as vmstat, netstat, iostat)

memory usage

CPU usage (especially important when client and server, or multiple clients, are on the same machine)

swapping activity

paging activity

network activity

I/O activity per controller
 
 

number of packets sent/received (use dbcc monitor)

CPU usage

cache hit rates

I/O activity

if you see deadlock, determine if it's in the data pages or the index pages of a non-clustered index

average packet size

look at the TDS packets

number of queries/traffic to network
 

 And, of course, from the application:

how long did the job run, or

what is the average response time

using gprof output can be very useful
 

  Step 5 - What Can Be Done

if the application allows it, try parallelizing batch operations (beware of increased possibility of causing deadlocks)

use named caches to prevent data that's going to be re-used from being flushed

look at hardware alternatives (if allowing write caching shows improvement in your development environment, then perhaps a solid state device might help at the production site)

check showplan output, verify that the optimizer is doing for you what you?d expect

look at the MRU/LRU strategy being used, override if necessary

are there table scans where an index should be used

set TCP_NODELAY on both the client and the server

test with various packet sizes to optimize network performance

if dynamic SQL is being used, are the statements being re-used?

do the work close to the data, use stored procedures

use temp tables

for non-shared data

for intermediate join results

to establish result sets without holding locks

they will be cleaned up automatically

use optimistic locking (via a timestamp or version number)

use the minimal level of locking for each command;
 

use isolation level 0 (dirty reads) wherever you canThe query analysis tools that you will use most often are:
 
 
 
 

3. Server

 It is possible to encounter deadlocks when many long-running transactions are executed at the same time in the same database. Deadlocks become more common as the lock contention increases between those transactions (decreasing concurrency).

 Well-designed applications can avoid deadlocks (but may not fully eliminate them) by always acquiring locks in the same order. Updates to multiple tables should always be performed in the same order, with no table updated twice.

 At all costs, you must code to avoid deadlocks BUT they will happen anyhow (you can influence the frequency). When a deadlock occurs, Adaptive Server Enterprise will terminate one of the transactions so that the other(s) can complete. Whenever this occurs, that transaction that was terminated must be started again - automatically if it occurs in an (unattended) batch process.

 Be consistent in the order that all transactions access the tables. For example, wherever possible, try to avoid a situation such as:
 
 

 
Transaction A	Transaction B
begin transaction  ...  update table_1  ...  update table_2  ...  end transaction	begin transaction  ...  update table_2  ...  update table_1  ...  end transaction

 
 
 
retry-on-deadlock logic should be placed in the application for multi-update transactions
 

When a deadlock occurs, Adaptive Server Enterprise :

chooses one of the connections as the victim (chooses the user whose process has accumulated the least amount of CPU time as the victim) and aborts the transaction, allowing the other connection to complete;

rolls back the transaction; and
returns an error code of 1205.
 

 The application needs to check the errorcode and automatically restart the transaction if and when a 1205 is returned.

 The following client pseudo-code illustrates retrying a transaction when it is chosen as a deadlock victim:  dobegin transaction

 update tab1 set fld1=@newval_1 where fld2=@testval_1

 update tab2 set fld1=@newval_2 where fld2=@testval_2  while (@@error != 1205)

 commit transaction  A word of caution - make sure your retry logic doesn?t put you into a continuous loop. Add a counter or some other check to allow the application to quit trying after multiple attempts.

update only the fields that need to be updated
 

 If you update all the fields in the record, the Adaptive Server Enterprise is going to have to lock all the indices, possibly unnecessarily. So especially avoid updating the key fields (when the values don?t change).

 strive for updates in place
 

Adaptive Server Enterprise 11 will perform the update "in place", that is just update the record on its current page as opposed to doing a delete and insert, as long as the record still fits on the page and the clustered index value hasn?t changed.

 If your records include variable length fields, you can use the max-records-per-page feature to spread your data, leaving enough room on the page to accommodate the expected expansion of these fields.

tracking the cause of deadlocks
 

(note: this following TechNote can be found in Answerbase)

 
 SYBASE TechNote

 Adaptive Server Enterprise Trace Flags for Tracking Deadlocks

 Summary

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

This note explains some Adaptive Server Enterprise trace flags which can be used to trace the cause of deadlocks.

 Content

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

If you encounter deadlocks in your application, start Adaptive Server Enterprise with trace flags -T3605 and -T1204 (not 1205). The output from these trace flags can be used to trace the cause of the deadlocks.
 
 

WARNING!

Use trace flag 1204 only to trace the cause of deadlocks as it can seriously degrade performance.
 
 

The following output illustrates how trace flag output can be used to trace the cause of deadlocks.
 
 

================== errorlog file ==================

 
*** Deadlock detected - process 16 trying to wait

on process 41

Deadlock chain -->

LOCK REQUEST INITIATING DEADLOCK: LOGICAL:

EX_PAGE at 0xaac690

lockid=107001 spid=16 dbid=6

Process 41 waiting on Process 16 for resource:

LOGICAL. Lock requested by

spid 41:

SH_PAGE at 0xac549c

lockid=106991 spid=41 dbid=6

BLOCKED by spid 16 with the following lock:

EX_PAGE Blocking at 0x20be7c0

lockid=106991 spid=16 dbid=6

pstat=0x0100 [ ]

VICTIM: process 41 ; pstat 0x0100 [ ] ; cputime = 26

Process 41 was selected as victim

Waking up victim process 41 (110 : 0x0100)

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

As this output illustrates, an exclusive lock is requested by

server process ID (spid) 41. Page 106,991 already has a

shared lock already held by spid 16. So, spid 41 must wait for

spid 16 to finish. On page 107,001, spid 16 requests an

exclusive lock. However, spid 41 already holds a lock on page

107,001 and deadlock occurs. To solve this problem, do the

following.
 
 

1. Issue the following isql commands:

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

1> select db_name(6)

2> go
 
 

1> use database_name

2> go
 
 

1> dbcc traceon(3604)

2> go
 
 

1> dbcc page(6, 107001) a

2> examine PAGE HEADER: objid, level and indid

3> go
 
 

1> select object_name( objid )

2> go
 
 

1> sp_helpindex table_name

2> go
 
 

2. Apply step 1 to page 106,991. In some cases, the

pages may have been deallocated.

 

3.2 Other Locking Issues

lock at the minimum isolation level required by the application
 

Adaptive Server Enterprise supports isolation levels 0 (uncommitted or ?dirty? reads), 1, and 3 (holdlocks). The locking level can be set at per command and/or per connection. Remember that whereas reads at isolation level 3 guarantee data consistency throughout the transaction, it also means that the locks are held until the end of the transaction. Other transactions will be blocked and have to wait until the transaction releases the locks causing poor performance.

short transactions


By keeping the transactions short, you lessen the time a lock will be held. Often some of the work might be done outside the transaction or a transaction might be broken into two or more smaller transactions.

For example, instead of: begin transaction
foreach deposit_record from deposit_table
update individual_accounts_table

delete deposit_record from deposit_table
end-foreach
commit transaction
 do this:
 foreach deposit_record from deposit_table
begin transaction
update individual_accounts_table
delete deposit_record from deposit_table
commit transaction
end-foreach

hold no locks during user ?think-time?
 

 If there is a lock held while data is displayed on a user?s terminal, waiting for him to decide on an action, what?s to prevent him from taking a coffee break, lunch break, or go on vacation? "But if there are no locks, how do I know that no one else has updated the record?"

One method to ensure data consistency is to have a timestamp on the record and use it to make sure that no one else has updated it.

 max-rows-per-page: spreading the data
  Lock contention on heavily accessed tables may be alleviated by spreading the data out using the max-rows-per-page feature. This feature can be used on the leaf-level pages on any index, clustered or non-clustered. (Of course, when you use it on the clustered index you?re spreading the actual data records.) There is a trade off here - while by spreading the data you will alleviate lock contention you?ll pay for this by using more disk, having to do more physical I/O, and requiring more memory for additional cache to hold the additional pages read.

avoid table locks (configurable lock promotion)
  By default a lock is put on the entire table once a command locks 200 pages in a table. This parameter (default value of 200) can be changed to a value greater than 200 but keep in mind that you may have to allocate more memory to handle the additional locks being held.

declare cursor?s intent (Read Only or Updateable)
 

 If you don?t explicitly make the declaration, then Adaptive Server Enterprise will most often choose to open the cursor as updateable which could lead to lock contention.

3.3 new Adaptive Server Enterprise 11.x Features to Boost Performance

named caches
 

Named Caches is a completely new feature for Adaptive Server Enterprise 11.x. No competing RDBMS product has anything of a similar concept, and it's ramifications are currently being evaluated outside Sybase to fully analyze the potential benefits.

 

Previous to Adaptive Server Enterprise 11.x, all data read in from disk resides in a single server-wide cache. When the cache becomes full, the next call for cache space is satisfied by discarding the contents of the least recently used³ cache pages and refilling them with the currently needed data. In a heavily loaded system those least-recently-used pages may well still be needed for other ongoing or repeated queries, which will consequently need to refetch them from disk. A single ill-timed ad-hoc 'select *' query from a large table might flush the cache of critical data that is currently or soon-to-be in use by multiple core business transactions. The Named Caches feature provides a means to avoid these costly repeated physical I/Os for regularly useful data. This feature allows the DBA to divide the available cache space into any number of named caches, with the remainder staying the default cache. The DBA can then bind one or more objects to a named cache. If the bound object(s) can fit entirely into the named cache, their pages will remain in memory once read in, and will be available for all queries but are insulated from flushing due to the cache needs of other queries. Note that queries which access huge uncacheable tables can still benefit greatly by binding the index(s) to a devoted cache. A table and index may be bound to separate caches. As an example, if a huge table was the target of occasional queries whose data weren't likely to be of repeated value in cache, the table might be bound to a purposely small named cache, enforcing that pages are swapped in-and-out as needed, never swamping even the default cache. Additionally, this table's index might be designed to cover most of the frequent queries, and could be bound separately to another cache which holds all of it. This is also true if the index is clustered, in that the non-leaf pages of a clustered index can be separately bound to a cache whether or not the data pages are.

Note that system tables can also be bound to a named cache. Sysindexes is typically a 'hot' table in OLTP environments. Lastly, note that you can bind a database to a named cache. Tempdb is a logical choice for OLTP using lots of temp tables.

Please note that some conservatism is warranted when applying this new capability because you still have a finite amount of cache to partition, and any cache space that is devoted to a named cache will be inaccessible for anything but the object(s) bound to the cache. The default cache behaves like the cache of previous releases, and should remain of a useful size for all the non-bound data.

 large I/Os
 

For some operations such as table scans and N-way joins, it would be possible for the optimizer to predict that the next N pages of a table are also going to be useful when it is asked to fetch page 1. This new feature allows the server to issue a single I/O call of up to 16k (8 pages) at a time. For any cache, including the default cache, the space of a cache can be divided into pools of different page sizes. The default configuration for any cache is a single pool of 2k pages.

 

The possible page sizes are 2k, 4k, 8k and 16k. Depending upon the object(s) bound to the cache and the intended use of the object(s) one might configure a 2k pool and a 16k pool. Note that the actual pool chosen for I/O depends on complexities such as whether the page is already in the cache in one or other of the pools. There is new syntax to force this prefetch behavior on or off. By default it is on. In some cases with prefetching multiple pages in a single I/O a table scan may become faster than an index scan. Here also, conservatism is apt. If circumstances force the use of the 2k pool for a query suited to the 16k pool, and the 2k pool is too small, the processing may have to wait while 2k pages are freed up for use.  If your application performs I/O on heaps, they should use a cache that allows large I/O. data slices
   Data slices is a feature aimed at improving the parallel performance of inserting rows to a table without a clustered index. It allows multiple parallel insertion processes to run without contention. Formerly these processes would contend for the last page of the heap, where each new row is inserted. The DBA is now able to specify a configurable number of slices to a table. Each of these slices has it's own 'last page' where the next row will be inserted. Thus for a 5-slice table, 5 inserters can run unimpeded.

Note that this is suited well for history, log tables and fast parallel BCP, but does not address the updating of indexes. If the insertions are in the order of any index the multiple processes will still serialize on the last page of that index.

 Limitations with Data Slices: A table cannot simultaneously have a clustered index and slices. Select statements going on in parallel with these inserts may incur/cause additional deadlocking because of the different order in which locks are obtained for reads and writes.

3.4 Server Helpful Hints

Simple Tricks to Use for Good Performance

 use ">=" rather than ">" whenever possible

"select * from foo where x > 3"

must scan all index pages where x=3 just to find the first qualified row!
"select * from foo where x >= 4"
can go directly to first qualified row.

example:  Using a table of 10,000 records with a non-unique, non-cluster index on fld_a (an integer). The distribution of the records is such that there are approximately 5% of the records (or 500) for each value of fld_a from 1 to 20.

query	results		logical reads	access method used
select count(*) from foo  where fld_a > 19	490	7		index scan beginning at fld_a = 19
select count(*) from foo  where fld_a >= 20	490	5		index scan beginning  at fld_a = 20
select count(*) from foo  where fld_a > 20	0	5		index scan beginning  at fld_a = 20
select count(*) from foo  where fld_a >= 21	0	2		index scan looking to begin at fld_a = 21
select count(*) from foo	10000	55		index scan beginning at first record of index
select count(*) from foo  where fld_x = fld_x	10000	625		scan of all data pages
select count(*) from foo  where fld_a >= 20  and fld_x = fld_x	490	495		index scan beginning  at fld_a = 20 but for  each row found the  datapage must be read

 

conclusions:    

The first 4 queries illustrate how using the ">=" instead of ">" save unnecessary I/O. The last 3 queries illustrate how covered queries reduce the I/O

use "EXISTS" and "IN" rather than "NOT EXISTS" and "NOT IN" (or COUNT)
 

Faster in both subqueries and IF statements Easy to re-write sprocs using EXISTS or IN. For example,  if not exists (select * from ...) begin

... /* statement group */

end could be re-written as :  if exists (select * from ...)

begin

goto exists_label

end

... /* statement group */

exists_label:

... EXISTS stop after 1st match as opposed to COUNT which does all the I/O to count!  example:  Using the same table of 10,000 records no index of fld_x, approximately 10% of the records have fld_x = 3.

query	results		logical reads
if not exists  (select * from foo  where fld_x = 3)	false	625
if exists  (select * from foo  where fld_x = 3)	true	1

 

conclusions:  A little restructuring of your logic flow can go a long way to improving performance.

IN clause
 

instead of just: where x in (@a, @b, @c)  add a between clause: where x in (@a, @b, @c)
and x between min(@a, @b, @c) and max(@a, @b, @c)  example:  Still using the same table of 10,000 records with the non-clustered, non-unique index of fld_a. conclusions:

query	results	logical reads
select count(*) from foo  where fld_a IN (8,12,13)	1468	55
select count(*) from foo  where fld_a IN (8,12,13) and fld_a between 8 and 13	1468	19

 

 At times, you can help the optimizer and therefore improve performance.

Avoid the following, if possible mathematical manipulation of SARGs (search arguments)
 

SELECT name FROM employee WHERE salary * 12 > 100000

 better to use

SELECT name FROM employee WHERE salary > 100000 / 12

example: Still using the same table of 10,000 records with the non-clustered, non-unique index of salary.

query	results	logical reads
select count(*) from foo  where salary * 12 > 72000	4990	75
select count(*) from foo  where salary > 72000/12	4990	39

 

conclusions:

 The optimizer is not very good at algebra, but if you simplify its task then the results will be improved performance.

use of Incompatible Datatypes between columns, SARGs, or SPROC Parameters
 

For example: float & int, char & varchar, binary & varbinary are incompatible. int and intn (allows nulls) are OK.

example:  Still using the same table of 10,000 records with a non-clustered, unique index on emp_id which is defined as numeric(8,0).

query	results	logical reads
select emp_id from foo  where emp_id =34	34	625
select emp_id from foo  where emp_id  =convert(numeric(8,0),34)	34	3

conclusions:

Experience has shown big wins when care is taken to prevent these type mismatches.

The following shows the output generated when the above 2 queries were ran with:

- set showplan on

- set statistics io on

- set statistics time on

- dbcc traceon(3604,302)

1> set statistics io on

2> set statistics time on

3> set showplan on

... 1> dbcc traceon(3604, 302)

...

1> select emp_id from foo

where emp_id = 34

Finishing q_score_index() for table 'foo' (objectid 16003088).

Cheapest index is index 4, costing 62 pages and generating 10000 rows per scan.

Index covers query.

Search argument selectivity is 1.000000.

*******************************

QUERY IS CONNECTED

0 -

NEW PLAN (total cost = 1240): varno=0 indexid=4 path=0x20707878 pathtype=sclause method=NESTED ITERATION

outerrows=1 rows=10000 joinsel=1 cpages=62 lp=62 pp=62 corder=1

TOTAL # PERMUTATIONS: 1

TOTAL # PLANS CONSIDERED: 2

FINAL PLAN (total cost = 1240):

 varno=0 indexid=4 path=0x20707878 pathtype=sclause method=NESTED ITERATION

outerrows=1 rows=10000 joinsel=1 cpages=62 lp=62 pp=62 corder=1

STEP 1

The type of query is SELECT.

FROM TABLE

foo

Nested iteration

Index : emp_emp_id

Parse and Compile Time 0.

Adaptive Server Enterprise cpu time: 0 ms.

emp_id

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

34.000000

Table: foo scan count 1, logical reads: 60, physical reads: 0

Total writes for this command: 0

Execution Time 0.

Adaptive Server Enterprise cpu time: 0 ms. SQL Server elapsed time: 243 ms.

(1 row affected) now using the convert function to ensure type matching

1> select emp_id from foo

where emp_id = convert(numeric(8,0),34)

*******************************

Entering q_score_index() for table 'foo' (objectid 16003088).

The table has 10000 rows and 625 pages.

Scoring the SEARCH CLAUSE: emp_id EQ

Base cost: indid: 0 rows: 10000 pages: 625

Unique nonclustered index found--return rows 1 pages 2

Cheapest index is index 4, costing 2 pages and generating 1 rows per scan.

Index covers query.

Search argument selectivity is 0.000100.

*******************************

QUERY IS CONNECTED

NEW PLAN FOR ONEROW (total cost = 40):

varno=0 indexid=4 path=0x2015386c pathtype=sclause method=NESTED ITERATION

outerrows=1 rows=1 joinsel=163854 cpages=2 lp=2 pp=2 corder=1

TOTAL # PERMUTATIONS: 0

TOTAL # PLANS CONSIDERED: 0

FINAL PLAN (total cost = 538558464):

varno=0 indexid=4 path=0x2015386c pathtype=sclause method=NESTED ITERATION

outerrows=1 rows=1 joinsel=163854 cpages=2 lp=2 pp=2 corder=1

STEP 1

The type of query is SELECT.

FROM TABLE

foo

Nested iteration

Index : emp_emp_id

Parse and Compile Time 0.

Adaptive Server Enterprise cpu time: 0 ms.

emp_id

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

34.000000

Table: foo scan count 1, logical reads: 2, physical reads: 0

Total writes for this command: 0

Execution Time 0.

Adaptive Server Enterprise cpu time: 0 ms. SQL Server elapsed time: 0 ms.

 (1 row affected)

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 scan the table! OR strategy requires additional cost of creating and sorting a work table. Evaluate UNIONs as an alternative !

not using the Leading Portion of the Index
 

(unless the query is completely covered)

use of Non-Equal Expressions (!=) in WHERE Clause
whenever possible, populate temporary table with "select into"
 

 SELECT INTO operations are not logged and therefore are significantly faster than using an INSERT with a nested SELECT.

3.5 trace flags with which to boot the Server
 
 
  1610 - TCP_NODELAY  (see section 6.1 below) 260 - DONE_IN_PROC  To facilitate SQLDebug, Adaptive Server Enterprise needs to append an additional 9 bytes of data (a "done_in_proc" message) to the TDS packet for every command in a Stored Procedure.

 These messages may impose additional TDS communications. Stored procedures with significant "server-based" processing can reduce this overhead by surpressing these "done_in_proc" messages.

 The default behaviour of the pre- Adaptive Server Enterprise 11.x Server was to send these 9 bytes. To prevent "done_in_proc" messages, start Adaptive Server Enterprise with trace flag #260. Note: your errorlog will not indicate the use of this option with any message!

 In the Adaptive Server Enterprise 11.x Server, the default behaviour has changed and more and easier control has been given to the user. A new configuration variable (called done_inproc) has been made available. The values for this variable are: 0 - "done in proc" turned off for the whole server
(default in Adaptive Server Enterprise 11.x)

1 - "done in proc" turned on for the whole server (default in pre-Adaptive Server Enterprise 11.x)

2 - "done in proc" turned on for this connection only

3 - "done in proc" turned off for this connection only Test Carefully, turning this off has been known to break some existing 3rd Party applications.