Re: playing with SDU/TDU

On Tue, 18 Dec 2001 16:44:21 GMT, Ed_Stevens_at_nospam.noway.nohow (Ed Stevens) wrote:

>Platform: Ora 8.0.5, NT4
>
>Do I have a problem with my testing methods or my expectations?
>
>SQL*Net related wait events continue to be my biggest hitters, so I decided to
>play around w/ SDU and TDU.
>
>First, I pulled the 5 most executed and 5 "most blocks read" queries from
>v$sqltext and put all in a sql script, adding a CONNECT at the top of the script
>and a DISCONNECT at the bottom.
>
>Next, set the SDU and TDU both to 32768 for the sid in question in LISTENER.ORA.
>
>Next, set the SDU and TDU both to 2048 for the sid in question in TNSNAMES.ORA.
>
>Next, run utlbstat, execute the script 20 times, then run utlestat.
>
>Repeat the 'bstat - execute - estat' cycle, incrementing SDU/TDU in tnsnames by
>2k increments until we reach a value of 20k.
>
>Examining the estat reports showed virtually no change in any of the SQL*Net
>wait events across the range of tests.
>
>Next I tried evaluating SQLNet traces. I set client_trace_level=16, then
>repeated the cycle of incrementing SDU/TDU in tnsnames from 2k to 20k in 2k
>increments. For these tests, I only executed the script once at each value, and
>isolated the trace files for each trial, then counted the total number of
>packets sent and received at each SDU value. Results are as follows:
>
>SDU value - packets received - packets sent
>2k 86 69
>4k 73 66
>6k 69 66
>8k 69 66
>10k 67 66
>12k 65 66
>14k 65 66
>16k 65 66
>18k 65 66
>20k 65 66
>
>All of these trials were run on a test server. The databases have approximately
>the same data volumes as production (they are refreshed from full production
>exports) but the user traffic is far less than in production.
>
>So, I see that increasing SDU resulted in a decrease of packets sent from the
>server to the client, until we reached SDU=12k, but there was no corresponding
>improvement in TNS* related wait events. So is this an indication that I should
>set my SDU/TDU values to 12k? Would that be a waste of time? Or am I on the
>right track but using flawed testing methods?

SDU should be a multiple of the MTU of the underlying protocol. On Windoze the MTU is 1500 or 1508 bytes.

Hth

Sybrand Bakker, Senior Oracle DBA

To reply remove -verwijderdit from my e-mail address Received on Tue Dec 18 2001 - 12:03:30 CST