Re: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Luck, Tony wrote: Only a new user would have to pull the whole history ... and for most uses it is sufficient to just pull the current top of the tree. Linus' own tree only has a history going back to 2.6.12.-rc2 (when he started using git). Someday there might be a server daemon that can batch up the changes for a "pull" to conserve network bandwidth. There is a mailing list "git@vger.kernel.org" where these issues are discussed. Archives are available at marc.theaimsgroup.com and gelato. Thanks tony I wasn't aware of the list, I'll look there for git info from now on. Best Regards, Stan -Tony - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
RE: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
>That said, is there any plan to change how this functions in the future >to solve these problems? I.e. have it not use so much diskspace and >thus use less bandwith. Am I misunderstanding in assuming that after >say 1000 commits go into the tree it could end up several megs or gigs >bigger? > >If that is the case might it not be more prudent to sort this out now? Only a new user would have to pull the whole history ... and for most uses it is sufficient to just pull the current top of the tree. Linus' own tree only has a history going back to 2.6.12.-rc2 (when he started using git). Someday there might be a server daemon that can batch up the changes for a "pull" to conserve network bandwidth. There is a mailing list "git@vger.kernel.org" where these issues are discussed. Archives are available at marc.theaimsgroup.com and gelato. -Tony - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Luck, Tony wrote: Yeah, I'm facing the same issue. I started playing with git last night. Apart from disk-space usage, it's very nice, though I really hope someone puts together a web-interface on top of git soon so we can seek what changed when and by whom. Disk space issues? A complete git repository of the Linux kernel with all changesets back to 2.4.0 takes just over 3G ... which is big compared to BK, but 3G of disk only costs about $1 (for IDE ... if you want 15K rpm SCSI, then you'll pay a lot more). Network bandwidth is likely to be a bigger problem. That said, is there any plan to change how this functions in the future to solve these problems? I.e. have it not use so much diskspace and thus use less bandwith. Am I misunderstanding in assuming that after say 1000 commits go into the tree it could end up several megs or gigs bigger? If that is the case might it not be more prudent to sort this out now? There's a prototype web i/f at http://grmso.net:8090/ that's already looking fairly slick. Yes it is very slick. Kudos to the creator. -sb -Tony - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
On Thu, 21 Apr 2005 10:33:29 -0700 David Mosberger wrote: | > On Thu, 21 Apr 2005 10:19:28 -0700, "Luck, Tony" <[EMAIL PROTECTED]> said: | | >> I just checked 2.6.12-rc3 and the fls() fix is indeed missing. | >> Do you know what happened? | | Tony> If BitKeeper were still in use, I'd have dropped that patch | Tony> into my "release" tree and asked Linus to "pull" ... but it's | Tony> not, and I was stalled. I should have a "git" tree up and | Tony> running in the next couple of days. I'll make sure that the | Tony> fls fix goes in early. | | Yeah, I'm facing the same issue. I started playing with git last | night. Apart from disk-space usage, it's very nice, though I really | hope someone puts together a web-interface on top of git soon so we | can seek what changed when and by whom. 2 people have already done that. Examples: http://ehlo.org/~kay/gitweb.pl and http://grmso.net:8090/ and the commits mailing list is now working. A script to show nightly (or daily:) commits and make a daily patch tarball is also close to ready. --- ~Randy - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
RE: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
> On Thu, 21 Apr 2005 10:41:52 -0700, "Luck, Tony" <[EMAIL PROTECTED]> said: Tony> Disk space issues? A complete git repository of the Linux Tony> kernel with all changesets back to 2.4.0 takes just over 3G Tony> ... which is big compared to BK, but 3G of disk only costs Tony> about $1 (for IDE ... if you want 15K rpm SCSI, then you'll Tony> pay a lot more). Network bandwidth is likely to be a bigger Tony> problem. Ever heard that data is a gas? My disks always fill up in no time at all, no matter how big they are. I agree that network bandwidth is an bigger issue, though. Tony> There's a prototype web i/f at http://grmso.net:8090/ that's Tony> already looking fairly slick. Indeed. Plus it has a cool name, too. Thanks for the pointer. --david - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
RE: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
>Yeah, I'm facing the same issue. I started playing with git last >night. Apart from disk-space usage, it's very nice, though I really >hope someone puts together a web-interface on top of git soon so we >can seek what changed when and by whom. Disk space issues? A complete git repository of the Linux kernel with all changesets back to 2.4.0 takes just over 3G ... which is big compared to BK, but 3G of disk only costs about $1 (for IDE ... if you want 15K rpm SCSI, then you'll pay a lot more). Network bandwidth is likely to be a bigger problem. There's a prototype web i/f at http://grmso.net:8090/ that's already looking fairly slick. -Tony - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
RE: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
> On Thu, 21 Apr 2005 10:19:28 -0700, "Luck, Tony" <[EMAIL PROTECTED]> said: >> I just checked 2.6.12-rc3 and the fls() fix is indeed missing. >> Do you know what happened? Tony> If BitKeeper were still in use, I'd have dropped that patch Tony> into my "release" tree and asked Linus to "pull" ... but it's Tony> not, and I was stalled. I should have a "git" tree up and Tony> running in the next couple of days. I'll make sure that the Tony> fls fix goes in early. Yeah, I'm facing the same issue. I started playing with git last night. Apart from disk-space usage, it's very nice, though I really hope someone puts together a web-interface on top of git soon so we can seek what changed when and by whom. --david - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
RE: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
>I just checked 2.6.12-rc3 and the fls() fix is indeed missing. Do you >know what happened? If BitKeeper were still in use, I'd have dropped that patch into my "release" tree and asked Linus to "pull" ... but it's not, and I was stalled. I should have a "git" tree up and running in the next couple of days. I'll make sure that the fls fix goes in early. -Tony - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [Gelato-technical] Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Tony and Andrew, I just checked 2.6.12-rc3 and the fls() fix is indeed missing. Do you know what happened? --david > On Thu, 21 Apr 2005 13:30:50 +0200, Andreas Hirstius <[EMAIL PROTECTED]> > said: Andreas> Hi, The fls() patch from David solves the problem :-)) Andreas> Do you have an idea, when it will be in the mainline Andreas> kernel?? Andreas> Andreas Andreas> Bartlomiej ZOLNIERKIEWICZ wrote: >> Hi! >> >>> A small update. >>> >>> Patching mm/filemap.c is not necessary in order to get the >>> improved performance! It's sufficient to remove >>> roundup_pow_of_two from |get_init_ra_size ... >>> >>> So a simple one-liner changes to picture dramatically. But why >>> ?!?!? >> >> >> roundup_pow_of_two() uses fls() and ia64 has buggy fls() >> implementation [ seems that David fixed it but patch is not in >> the mainline yet]: >> >> http://www.mail-archive.com/linux-ia64@vger.kernel.org/msg01196.html >> >> That would also explain why you couldn't reproduce the problem on >> ia32 Xeon machines. >> >> Bartlomiej >> Andreas> ___ Andreas> Gelato-technical mailing list Andreas> [EMAIL PROTECTED] Andreas> https://www.gelato.unsw.edu.au/mailman/listinfo/gelato-technical - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Hi, The fls() patch from David solves the problem :-)) Do you have an idea, when it will be in the mainline kernel?? Andreas Bartlomiej ZOLNIERKIEWICZ wrote: Hi! A small update. Patching mm/filemap.c is not necessary in order to get the improved performance! It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ... So a simple one-liner changes to picture dramatically. But why ?!?!? roundup_pow_of_two() uses fls() and ia64 has buggy fls() implementation [ seems that David fixed it but patch is not in the mainline yet]: http://www.mail-archive.com/linux-ia64@vger.kernel.org/msg01196.html That would also explain why you couldn't reproduce the problem on ia32 Xeon machines. Bartlomiej - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Hi! A small update. Patching mm/filemap.c is not necessary in order to get the improved performance! It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ... So a simple one-liner changes to picture dramatically. But why ?!?!? roundup_pow_of_two() uses fls() and ia64 has buggy fls() implementation [ seems that David fixed it but patch is not in the mainline yet]: http://www.mail-archive.com/linux-ia64@vger.kernel.org/msg01196.html That would also explain why you couldn't reproduce the problem on ia32 Xeon machines. Bartlomiej - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
A small update. Patching mm/filemap.c is not necessary in order to get the improved performance! It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ... So a simple one-liner changes to picture dramatically. But why ?!?!? Andreas | jmerkey wrote: For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved". /linux/include/linux/blkdev.h //#define BLKDEV_MIN_RQ 4 //#define BLKDEV_MAX_RQ 128 /* Default maximum */ #define BLKDEV_MIN_RQ 4096 #define BLKDEV_MAX_RQ 8192 /* Default maximum */ Jeff Andreas Hirstius wrote: Hi, We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit). The config for the kernels are basically identical. Here is some vmstat output: kernel 2.6.9: ~1GB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74 kernel 2.6.11: ~55MB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75 Because the filesystem might have an impact on the measurement, "dd" on /dev/md0 was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size. For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated: 1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps) 2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate 3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate) 4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps) 5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages) I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!! With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size. This simple patch solves the problem, but I have no idea of possible side-effects ... --- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.0 +0200 @@ -719,7 +719,7 @@ index = *ppos >> PAGE_CACHE_SHIFT; next_index = index; prev_index = ra.prev_page; - last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT; + last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> PAGE_CACHE_SHIFT; offset = *ppos & ~PAGE_CACHE_MASK; isize = i_size_read(inode); --- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.0 +0200 @@ -70,7 +70,7 @@ */ static unsigned long get_init_ra_size(unsigned long size, unsigned long max) { - unsigned long newsize = roundup_pow_of_two(size); + unsigned long newsize = size; if (newsize <= max / 64) newsize = newsize * newsize; In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots: http://www.cern.ch/openlab-debugging/raid Regards, Andreas Hirstius - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
On Wed, 2005-04-20 at 10:55 -0600, jmerkey wrote: > > For 3Ware, you need to chage the queue depths, and you will see > dramatically improved performance. 3Ware can take requests > a lot faster than Linux pushes them out. Try changing this instead, you > won't be going to sleep all the time waiting on the read/write > request queues to get "unstarved". > > > /linux/include/linux/blkdev.h > > //#define BLKDEV_MIN_RQ 4 > //#define BLKDEV_MAX_RQ 128 /* Default maximum */ > #define BLKDEV_MIN_RQ 4096 > #define BLKDEV_MAX_RQ 8192 /* Default maximum */ > BTW, don't do this. BLKDEV_MIN_RQ sets the size of the mempool reserved requests and will only get slightly used in low memory conditions, so most memory will probably be wasted. Just change /sys/block/xxx/queue/nr_requests Nick - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Burst is good. There's another window in the SCSI layer that limits to bursts of 128 sector runs (this seems to be the behavior on 3Ware). I've never changed this, but increasing the max number of SCSI requests at this layer may help. The bursty behavior is good, BTW. Jeff Andreas Hirstius wrote: I was curious if your patch would change the write rate because I see only ~550MB/s (continuous) which is about a factor two away from the capabilities of the disks. ... and got this behaviour (with and without my other patch): (with single "dd if=/dev/zero of=testxx bs=65536 count=15 &" or several of them in parallel on an XFS fs) "vmstat 1" output 0 0 0 28416 37888 1577836800 0 0 8485 3043 0 0 0 100 6 0 0 22144 37952 1578592000 0 12356 7695 2029 0 61 0 39 7 0 0 20864 38016 1578585600 324 1722240 8046 4159 0 100 0 0 7 0 0 20864 38016 1578476800 0 1261440 8391 5222 0 100 0 0 7 0 0 25984 38016 1578150400 0 2003456 8372 5038 0 100 0 0 0 6 0 22784 38016 1578150400 0 2826624 8397 8423 0 93 7 0 0 0 0 21632 38016 1578368000 0 147840 8572 12114 0 9 17 74 0 0 0 21632 38016 1578368000 052 8586 5185 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8588 5412 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8580 5372 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 7840 5590 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8587 5321 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8569 5575 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8550 5157 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 7963 5640 0 0 0 100 0 0 0 21632 38016 1578368000 032 8583 4434 0 0 0 100 7 0 0 20800 38016 1578476800 0 7424 8404 3638 0 15 0 85 8 0 0 20864 38016 1578694400 0 688768 7357 3221 0 100 0 0 8 0 0 20736 28544 1579424000 0 1978560 8376 4897 0 100 0 0 7 0 0 22208 20736 1579878400 0 1385088 8367 4984 0 100 0 0 6 0 0 22144 6848 158126720056 1291904 8377 4815 0 100 0 0 0 0 0 50240 6848 1580940800 304 3136 8556 5088 1 26 0 74 0 0 0 50304 6848 1580940800 0 0 8572 5181 0 0 0 100 The average rate here is again pretty close to 550MB/s, it just writes the blocks in "bursts"... Andreas jmerkey wrote: For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved". /linux/include/linux/blkdev.h //#define BLKDEV_MIN_RQ 4 //#define BLKDEV_MAX_RQ 128 /* Default maximum */ #define BLKDEV_MIN_RQ 4096 #define BLKDEV_MAX_RQ 8192 /* Default maximum */ Jeff Andreas Hirstius wrote: Hi, We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit). The config for the kernels are basically identical. Here is some vmstat output: kernel 2.6.9: ~1GB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74 kernel 2.6.11: ~55MB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75 Because the filesystem might have an impact on the measurement, "dd" on /dev/md0 was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size. For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated: 1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps) 2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. r
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
I was curious if your patch would change the write rate because I see only ~550MB/s (continuous) which is about a factor two away from the capabilities of the disks. ... and got this behaviour (with and without my other patch): (with single "dd if=/dev/zero of=testxx bs=65536 count=15 &" or several of them in parallel on an XFS fs) "vmstat 1" output 0 0 0 28416 37888 1577836800 0 0 8485 3043 0 0 0 100 6 0 0 22144 37952 1578592000 0 12356 7695 2029 0 61 0 39 7 0 0 20864 38016 1578585600 324 1722240 8046 4159 0 100 0 0 7 0 0 20864 38016 1578476800 0 1261440 8391 5222 0 100 0 0 7 0 0 25984 38016 1578150400 0 2003456 8372 5038 0 100 0 0 0 6 0 22784 38016 1578150400 0 2826624 8397 8423 0 93 7 0 0 0 0 21632 38016 1578368000 0 147840 8572 12114 0 9 17 74 0 0 0 21632 38016 1578368000 052 8586 5185 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8588 5412 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8580 5372 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 7840 5590 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8587 5321 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8569 5575 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 8550 5157 0 0 0 100 0 0 0 21632 38016 1578368000 0 0 7963 5640 0 0 0 100 0 0 0 21632 38016 1578368000 032 8583 4434 0 0 0 100 7 0 0 20800 38016 1578476800 0 7424 8404 3638 0 15 0 85 8 0 0 20864 38016 1578694400 0 688768 7357 3221 0 100 0 0 8 0 0 20736 28544 1579424000 0 1978560 8376 4897 0 100 0 0 7 0 0 22208 20736 1579878400 0 1385088 8367 4984 0 100 0 0 6 0 0 22144 6848 158126720056 1291904 8377 4815 0 100 0 0 0 0 0 50240 6848 1580940800 304 3136 8556 5088 1 26 0 74 0 0 0 50304 6848 1580940800 0 0 8572 5181 0 0 0 100 The average rate here is again pretty close to 550MB/s, it just writes the blocks in "bursts"... Andreas jmerkey wrote: For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved". /linux/include/linux/blkdev.h //#define BLKDEV_MIN_RQ 4 //#define BLKDEV_MAX_RQ 128 /* Default maximum */ #define BLKDEV_MIN_RQ 4096 #define BLKDEV_MAX_RQ 8192 /* Default maximum */ Jeff Andreas Hirstius wrote: Hi, We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit). The config for the kernels are basically identical. Here is some vmstat output: kernel 2.6.9: ~1GB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74 kernel 2.6.11: ~55MB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75 Because the filesystem might have an impact on the measurement, "dd" on /dev/md0 was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size. For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated: 1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps) 2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate 3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate) 4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps) 5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k p
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Just tried it, but the performance problem remains :-( (actually, why should it change? This part of the code didn't change so much between 2.6.10-bk6 and -bk7...) Andreas jmerkey wrote: For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved". /linux/include/linux/blkdev.h //#define BLKDEV_MIN_RQ 4 //#define BLKDEV_MAX_RQ 128 /* Default maximum */ #define BLKDEV_MIN_RQ 4096 #define BLKDEV_MAX_RQ 8192 /* Default maximum */ Jeff Andreas Hirstius wrote: Hi, We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit). The config for the kernels are basically identical. Here is some vmstat output: kernel 2.6.9: ~1GB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74 kernel 2.6.11: ~55MB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75 Because the filesystem might have an impact on the measurement, "dd" on /dev/md0 was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size. For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated: 1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps) 2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate 3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate) 4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps) 5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages) I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!! With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size. This simple patch solves the problem, but I have no idea of possible side-effects ... --- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.0 +0200 @@ -719,7 +719,7 @@ index = *ppos >> PAGE_CACHE_SHIFT; next_index = index; prev_index = ra.prev_page; - last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT; + last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> PAGE_CACHE_SHIFT; offset = *ppos & ~PAGE_CACHE_MASK; isize = i_size_read(inode); --- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.0 +0200 @@ -70,7 +70,7 @@ */ static unsigned long get_init_ra_size(unsigned long size, unsigned long max) { - unsigned long newsize = roundup_pow_of_two(size); + unsigned long newsize = size; if (newsize <= max / 64) newsize = newsize * newsize; In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots: http://www.cern.ch/openlab-debugging/raid Regards, Andreas Hirstius - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved". /linux/include/linux/blkdev.h //#define BLKDEV_MIN_RQ 4 //#define BLKDEV_MAX_RQ 128 /* Default maximum */ #define BLKDEV_MIN_RQ 4096 #define BLKDEV_MAX_RQ 8192 /* Default maximum */ Jeff Andreas Hirstius wrote: Hi, We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit). The config for the kernels are basically identical. Here is some vmstat output: kernel 2.6.9: ~1GB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74 kernel 2.6.11: ~55MB/s read procs memory swap io system cpu r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75 Because the filesystem might have an impact on the measurement, "dd" on /dev/md0 was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size. For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated: 1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps) 2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate 3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate) 4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps) 5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages) I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!! With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size. This simple patch solves the problem, but I have no idea of possible side-effects ... --- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.0 +0200 @@ -719,7 +719,7 @@ index = *ppos >> PAGE_CACHE_SHIFT; next_index = index; prev_index = ra.prev_page; - last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT; + last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> PAGE_CACHE_SHIFT; offset = *ppos & ~PAGE_CACHE_MASK; isize = i_size_read(inode); --- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.0 +0200 @@ -70,7 +70,7 @@ */ static unsigned long get_init_ra_size(unsigned long size, unsigned long max) { - unsigned long newsize = roundup_pow_of_two(size); + unsigned long newsize = size; if (newsize <= max / 64) newsize = newsize * newsize; In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots: http://www.cern.ch/openlab-debugging/raid Regards, Andreas Hirstius - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later
Hi, We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit). The config for the kernels are basically identical. Here is some vmstat output: kernel 2.6.9: ~1GB/s read procs memory swap io system cpu r b swpd free buff cache si sobibo incs us sy wa id 1 1 0 12672 6592 1591411200 108134456 15719 1583 0 11 14 74 1 0 0 12672 6592 1591520000 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 1591411200 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 1591411200 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 1591411200 1130496 0 16006 1586 0 12 14 74 kernel 2.6.11: ~55MB/s read procs memory swap io system cpu r b swpd free buff cache si sobibo incs us sy wa id 1 1 0 24448 37568 1590598400 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 1590924800 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 1590707200 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 1590380800 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 1590816000 57267 0 5168 1871 0 1 24 75 Because the filesystem might have an impact on the measurement, "dd" on /dev/md0 was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size. For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated: 1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps) 2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate 3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate) 4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps) 5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages) I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!! With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size. This simple patch solves the problem, but I have no idea of possible side-effects ... --- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.0 +0200 @@ -719,7 +719,7 @@ index = *ppos >> PAGE_CACHE_SHIFT; next_index = index; prev_index = ra.prev_page; - last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT; + last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> PAGE_CACHE_SHIFT; offset = *ppos & ~PAGE_CACHE_MASK; isize = i_size_read(inode); --- linux-2.6.12-rc2_orig/mm/readahead.c2005-04-04 18:40:05.0 +0200 +++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.0 +0200 @@ -70,7 +70,7 @@ */ static unsigned long get_init_ra_size(unsigned long size, unsigned long max) { - unsigned long newsize = roundup_pow_of_two(size); + unsigned long newsize = size; if (newsize <= max / 64) newsize = newsize * newsize; In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots: http://www.cern.ch/openlab-debugging/raid Regards, Andreas Hirstius - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/