----- Original Message ----- From: "Shawn" <[EMAIL PROTECTED]> To: "gentoo.org ML" <[EMAIL PROTECTED]> Sent: Saturday, December 20, 2003 5:13 AM Subject: Re: [gentoo-user] Kernel 2.6 test must read !!!!
>Did you ever think CPU accounting might be slightly different? Well I'm not sure what are you trying to say now, because you don't refer to a certain part of my post, it would certainly help if you would refer to the part in text you have doubts about. Maybe another explanation helps, top only gets confused when nptl is compiled into glibc, if not top is fine and shows correct values. So what CPU accounting are you talking about ??? The procps version I updated to is supposed to support nptl(so top should support it), however that is not the issue here(only a sideeffect), the issue is the cpu time used by overnet and without nptl I can be absolutely sure, that my programms show the correct cpu time values. Since I ran tests with overnet over the last few weeks with kernel 2.4 to determine the effects, that kerneltuning has on the tcp performance of overnet, I know exactly how much cpu time overnet uses in a specific state. That's why I ask for someone with a webserver or database server to run a test in case he has old statistics with kernel 2.4. I will continue to run a few tests and try to locate the cause for this issue, after I reemerged my apps wthout nptl again. >Unless you know how to really profile the app I'd defer to the wisdom of >others. Sounds like you think a syntetic test is worth more than a real life performance test. Kernel 2.6 is supposed to be faster, but did a lot of people prove it in real world yet ???? On Fri, 2003-12-19 at 18:16, SN wrote: > Okay, I want to summarize my experience with kernel 2.6 now. > > I upgraded to the 2.6 final yesterday and ran some tests yesterday and all > day today. > First of all, kernel 2.6 worked fine, no crashes , all my drivers worked all > my apps worked, all my hardware worked. > > Test with overnet however revealed some issues, overnets CPU time usage is > much higher with kernel 2.6. > > The first test I ran was overnet with glibc without nptl, overnet used about > twice as much cpu time as before with kernel 2.4. > For the second test I installed linux-headers 2.6-beta11 and compiled glibc > with nptl (use flag nptl was set), the result stayed the same. > > So overnet used about 15-20% cpu time with kernel 2.6 , with 2.4 it was > around 5-10% . > > To be able to compare the two results I ran the tests for hours and compared > the cpu time whenever overnet build up aproximately the same number of tcp > connections(measured with netstat) and the same number of openfiles(measured > with lsof). > Another effect: > After compiling in nptl I noticed, that top gets confused by nptl the values > it shows are not correct anymore. For example while compiling the > accumulated cpu time is 90% but the compile process cc1 shows that it is > using 2-5% cpu time and all other processes show about 0%, cc1 normally is > around 90% while compiling, I upgraded procps to the latest release because > top belongs to procps, but that didn't help it. So top is not working > corectly with nptl. > > > Although the system seems very responsive now I think something is wrong > with 2.6 > I'm not really sure what causes this, it could be due to the new TCP Stack, > a result of the new sheduling, or thread handling. Who knows??? > > I would love to see a discussion about that and maybe some other tests with > other applications, for example a webserver or database server under high > load. > > GrÃÃe Stefan > > > > > > > -- > [EMAIL PROTECTED] mailing list > -- [EMAIL PROTECTED] mailing list -- [EMAIL PROTECTED] mailing list
