Re: [gentoo-user] IO latency issues
maybe it will be fixed in kernel 3.9 http://www.phoronix.com/scan.php?page=news_itempx=MTMxNzA 2013/3/11 Florian Philipp li...@binarywings.net Am 10.03.2013 13:48, schrieb Florian Philipp: Am 10.03.2013 00:53, schrieb cosmoslx lin: 2013/3/10 Volker Armin Hemmann volkerar...@googlemail.com mailto:volkerar...@googlemail.com Am 09.03.2013 19:15, schrieb Florian Philipp: Hi list! Whenever I do sequential IO for a long stretch of time (e.g. md5summing 40 GB), I'm experiencing high load (ca. 6 on a 4 CPU system) and temporary freezes of most applications. For example, switching between tabs in konsole sometimes takes more than 2 seconds. [...] congratulation. You hit 'the bug'. Been around for ages, but for magical reasons kernel dev are unable to see or unable to do something about it. If you are using a vanilla kernel, posting on lkml might be the right thing to do. I have try the BFQ patch outside of the kernel mainline, it works well. Maybe you would like to see: http://algo.ing.unimo.it/people/paolo/disk_sched/ Thanks. I'll try pf-sources which should to include that patch set. Regards, Florian Philipp pf-sources (only BFQ enabled, not BFS) helped somewhat. kswapd still likes to keep one CPU busy when IO is happening but I suspect cleancache to be the culprit for that. I'll test without it later. Regards, Florian Philipp -- Best regards! Yu-yu Lin(林育宇) The Guangdong Key Laboratory of Information Security Technology (IST), School of Information Science and Technology, Sun Yat-sen (Zhongshan) University (中山大学), Guangzhou,P.R.China. Email: cosmo...@gmail.com cosmo...@gmail.com
Re: [gentoo-user] IO latency issues
Am 10.03.2013 00:53, schrieb cosmoslx lin: 2013/3/10 Volker Armin Hemmann volkerar...@googlemail.com mailto:volkerar...@googlemail.com Am 09.03.2013 19:15, schrieb Florian Philipp: Hi list! Whenever I do sequential IO for a long stretch of time (e.g. md5summing 40 GB), I'm experiencing high load (ca. 6 on a 4 CPU system) and temporary freezes of most applications. For example, switching between tabs in konsole sometimes takes more than 2 seconds. [...] congratulation. You hit 'the bug'. Been around for ages, but for magical reasons kernel dev are unable to see or unable to do something about it. If you are using a vanilla kernel, posting on lkml might be the right thing to do. I have try the BFQ patch outside of the kernel mainline, it works well. Maybe you would like to see: http://algo.ing.unimo.it/people/paolo/disk_sched/ Thanks. I'll try pf-sources which should to include that patch set. Regards, Florian Philipp signature.asc Description: OpenPGP digital signature
Re: [gentoo-user] IO latency issues
Am 10.03.2013 13:48, schrieb Florian Philipp: Am 10.03.2013 00:53, schrieb cosmoslx lin: 2013/3/10 Volker Armin Hemmann volkerar...@googlemail.com mailto:volkerar...@googlemail.com Am 09.03.2013 19:15, schrieb Florian Philipp: Hi list! Whenever I do sequential IO for a long stretch of time (e.g. md5summing 40 GB), I'm experiencing high load (ca. 6 on a 4 CPU system) and temporary freezes of most applications. For example, switching between tabs in konsole sometimes takes more than 2 seconds. [...] congratulation. You hit 'the bug'. Been around for ages, but for magical reasons kernel dev are unable to see or unable to do something about it. If you are using a vanilla kernel, posting on lkml might be the right thing to do. I have try the BFQ patch outside of the kernel mainline, it works well. Maybe you would like to see: http://algo.ing.unimo.it/people/paolo/disk_sched/ Thanks. I'll try pf-sources which should to include that patch set. Regards, Florian Philipp pf-sources (only BFQ enabled, not BFS) helped somewhat. kswapd still likes to keep one CPU busy when IO is happening but I suspect cleancache to be the culprit for that. I'll test without it later. Regards, Florian Philipp signature.asc Description: OpenPGP digital signature
Re: [gentoo-user] IO latency issues
Am 09.03.2013 19:15, schrieb Florian Philipp: Hi list! Whenever I do sequential IO for a long stretch of time (e.g. md5summing 40 GB), I'm experiencing high load (ca. 6 on a 4 CPU system) and temporary freezes of most applications. For example, switching between tabs in konsole sometimes takes more than 2 seconds. When doing this on an ext4 filesystem, the load seems to result from khugepaged and kswapd0 as well as some kworkers. I think I've had similar issues with NFS over wifi but I cannot test this now. Today I copied 60GB from my hard disk to an USB disk formatted with NTFS, issuing the copy command from KDE's dolphin. The freezes became so long it was impossible to work and then X11 locked up and had to be killed. I tried using a preemptive kernel but that didn't seem to help. blkio and cpu cgroups didn't help either. Ionice seems to be the only solution but while I'm okay with that, my dad won't be. Can anyone tell me what is causing this behavior? Throughput is good, by the way. That's why I don't suspect a driver issue. Thanks in advance! Florian Philipp congratulation. You hit 'the bug'. Been around for ages, but for magical reasons kernel dev are unable to see or unable to do something about it. If you are using a vanilla kernel, posting on lkml might be the right thing to do.
Re: [gentoo-user] IO latency issues
I have try the BFQ patch outside of the kernel mainline, it works well. Maybe you would like to see: http://algo.ing.unimo.it/people/paolo/disk_sched/ Also there is a 15-minute demo of the performance of BFQ: http://youtu.be/J-e7LnJblm8 2013/3/10 Volker Armin Hemmann volkerar...@googlemail.com Am 09.03.2013 19:15, schrieb Florian Philipp: Hi list! Whenever I do sequential IO for a long stretch of time (e.g. md5summing 40 GB), I'm experiencing high load (ca. 6 on a 4 CPU system) and temporary freezes of most applications. For example, switching between tabs in konsole sometimes takes more than 2 seconds. When doing this on an ext4 filesystem, the load seems to result from khugepaged and kswapd0 as well as some kworkers. I think I've had similar issues with NFS over wifi but I cannot test this now. Today I copied 60GB from my hard disk to an USB disk formatted with NTFS, issuing the copy command from KDE's dolphin. The freezes became so long it was impossible to work and then X11 locked up and had to be killed. I tried using a preemptive kernel but that didn't seem to help. blkio and cpu cgroups didn't help either. Ionice seems to be the only solution but while I'm okay with that, my dad won't be. Can anyone tell me what is causing this behavior? Throughput is good, by the way. That's why I don't suspect a driver issue. Thanks in advance! Florian Philipp congratulation. You hit 'the bug'. Been around for ages, but for magical reasons kernel dev are unable to see or unable to do something about it. If you are using a vanilla kernel, posting on lkml might be the right thing to do. -- Best regards! Yu-yu Lin(林育宇) The Guangdong Key Laboratory of Information Security Technology (IST), School of Information Science and Technology, Sun Yat-sen (Zhongshan) University (中山大学), Guangzhou,P.R.China. Email: cosmo...@gmail.com cosmo...@gmail.com
