Re[2]: [PATCH 07/11] md: rewrite handle_stripe_dirtying6 in asynchronous way
On Friday, January 16, 2009 you wrote: > On Thu, Jan 15, 2009 at 2:51 PM, Dan Williams > wrote: >> On Mon, Dec 8, 2008 at 2:57 PM, Yuri Tikhonov wrote: >> What's the reasoning behind changing the logic here, i.e. removing >> must_compute and such? I'd feel more comfortable seeing copy and >> paste where possible with cleanups separated out into their own patch. >> > Ok, I now see why this change was made. Please make this changelog > more descriptive than "Rewrite handle_stripe_dirtying6 function to > work asynchronously." Sure, how about the following: " md: rewrite handle_stripe_dirtying6 in asynchronous way Processing stripe dirtying in asynchronous way requires some changes to the handle_stripe_dirtying6() algorithm. In the synchronous implementation of the stripe dirtying we processed dirtying of a degraded stripe (with partially changed strip(s) located on the failed drive(s)) inside one handle_stripe_dirtying6() call: - we computed the missed strips from the old parities, and thus got the fully up-to-date stripe, then - we did reconstruction using the new data to write. In the asynchronous case of handle_stripe_dirtying6() we don't process anything right inside this function (since we under the lock), but only schedule the necessary operations with flags. Thus, if handle_stripe_dirtying6() is performed on the top of a degraded array we should schedule the reconstruction operation when the failed strips are marked (by previously called fetch_block6()) as to be computed (with the R5_Wantcompute flag), and all the other strips of the stripe are UPTODATE. The schedule_reconstruction() function will set the STRIPE_OP_POSTXOR flag [for new parity calculation], which is then handled in raid_run_ops() after the STRIPE_OP_COMPUTE_BLK one [which causes computing of the data missed]. " Regards, Yuri -- Yuri Tikhonov, Senior Software Engineer Emcraft Systems, www.emcraft.com ___ Linuxppc-dev mailing list Linuxppc-dev@ozlabs.org https://ozlabs.org/mailman/listinfo/linuxppc-dev
Re[2]: [PATCH 07/11] md: rewrite handle_stripe_dirtying6 in asynchronous way
Hello Cheng, On Friday, January 16, 2009 you wrote: > Ack, could you please make the changelog more descriptive? > and or add some of your benchmark results? Of course. We did benchmarking using the Xdd tool like follows: # xdd -op write -kbytes $kbytes -reqsize $reqsize -dio-passes 2 –verbose -target $target_device where $kbytes = data disks * size of disk $reqsize= data disks * chunk size $target_device = /dev/md0 This way we did write of full array size, and thus achieved the maximum performance. The test cases were RAID-6 built on the top of 14 S-ATA drives connected to 2 LSI cards (7+7) inserted into the 800 MHz Katmai board (based on ppc440spe) equipped with 4GB of 800 MHz DRAM . Here are the results (Psw - write throughput with s/w RAID-6; Phw - write throughput with the h/w accelerated RAID-6): PAGE_SIZE=4KB, chunk=64/128/256 KB Psw = 71/72/74 MBps Phw = 128/136/139 MBps PAGE_SIZE=16KB, chunk=256/512/1024 KB Psw = 81/81/82 MBps Phw = 205/244/239 MBps PAGE_SIZE=64KB, chunk=1024/2048/4096 KB Psw = 84/84/85 MBps Phw = 258/253/258 MBps PAGE_SIZE=256KB, chunk=4096/8192/16384 KB Psw = 81/83/83 MBps Phw = 288/275/274 MBps Regards, Yuri -- Yuri Tikhonov, Senior Software Engineer Emcraft Systems, www.emcraft.com ___ Linuxppc-dev mailing list Linuxppc-dev@ozlabs.org https://ozlabs.org/mailman/listinfo/linuxppc-dev
Re: Re[2]: [PATCH 07/11] md: rewrite handle_stripe_dirtying6 in asynchronous way
On Fri, Jan 16, 2009 at 7:24 AM, Yuri Tikhonov wrote: >> Ok, I now see why this change was made. Please make this changelog >> more descriptive than "Rewrite handle_stripe_dirtying6 function to >> work asynchronously." > > Sure, how about the following: > > " > > md: rewrite handle_stripe_dirtying6 in asynchronous way > > Processing stripe dirtying in asynchronous way requires some changes > to the handle_stripe_dirtying6() algorithm. > > In the synchronous implementation of the stripe dirtying we processed > dirtying of a degraded stripe (with partially changed strip(s) located > on the failed drive(s)) inside one handle_stripe_dirtying6() call: > - we computed the missed strips from the old parities, and thus got > the fully up-to-date stripe, then > - we did reconstruction using the new data to write. > > In the asynchronous case of handle_stripe_dirtying6() we don't > process anything right inside this function (since we under the lock), > but only schedule the necessary operations with flags. Thus, if > handle_stripe_dirtying6() is performed on the top of a degraded array > we should schedule the reconstruction operation when the failed strips > are marked (by previously called fetch_block6()) as to be computed > (with the R5_Wantcompute flag), and all the other strips of the stripe > are UPTODATE. The schedule_reconstruction() function will set the > STRIPE_OP_POSTXOR flag [for new parity calculation], which is then > handled in raid_run_ops() after the STRIPE_OP_COMPUTE_BLK one [which > causes computing of the data missed]. > > " Excellent! Thanks, Dan ___ Linuxppc-dev mailing list Linuxppc-dev@ozlabs.org https://ozlabs.org/mailman/listinfo/linuxppc-dev