Aaron Bannert wrote: > On Sat, Dec 22, 2001 at 02:02:53PM +0100, Sander Striker wrote: > >>>From: David Reid [mailto:[EMAIL PROTECTED]] >>>Sent: 22 December 2001 13:57 >>> >>>The transaction stats were what jumped out at me - 7% increase in failed >>>connections doesn't sound good to me :( But, then maybe I'm reading that >>>wrong? >>> >>Which is what I saw first too. But when I talked to Ian over irc <snippet>: >> >>[00:22] <IanHolsman> hey.. did you see the benchmark.. >>[00:22] <IanHolsman> could you parse it >>[00:22] <sander> Yes, saw it. >>[00:22] <sander> No, could not parse. >>[00:22] <IanHolsman> aah >>[00:23] <IanHolsman> http://webperf.org/a2/caw/29/Current Total HTTP and TCP Errors >vs Load 21-Dec-2001 1215.gif is probably the >>best image >>[00:24] <sander> Under a higher load we get more errors with v30? >>[00:24] <IanHolsman> if you look on v29 errors start happening around 500 users >sessions. with v30 they happen at 700 >>[00:24] <IanHolsman> no... under a higher load you get less. >>[00:24] <sander> Ah >> > > How are "Unsuccessful" transactions defined? Failed connect()ions, incomplete > HTTP requests? Invalid data returned? > > -aaron > >
I think I forgot to mention that there was a core dump in .30. as far as what a incomplete transaction is, I don't know. (I'll dig up the books) the test was a page fetch of a random page which had 7-10 includes. The # of sessions was increased every minute so that by the end of the test we had 1,000 open connections. we were getting ~100 connection/second for a lot of the time, but the response time & failure rate was getting worse as the number of sessions increased. This is probably better seen if I expand the CPU load RRD graph so it is wider and only for the first 15 minutes. both v30 & v29 couldn't handle the load at the end of the test, the major difference IMHO was v29 bailed out much sooner than v30 v30 also had a core dump in a weird location, which points to a memory corruption http://webperf.org/a2/caw/cpu_comp.html shows a comparision of server load, and transactions per second. the chart is a bit funky underneath, but what it shows is that v30 experienced a much lower error rate and higher TPS than v29 did. The next set of tests I do I will vary the Tranasactions per second. as I find that easier to understand and explain to people.
