Raghavendra Gowdappa
2018-Aug-30 15:18 UTC
[Gluster-users] [External] Re: file metadata operations performance - gluster 4.1
On Thu, Aug 30, 2018 at 8:38 PM, Davide Obbi <davide.obbi at booking.com> wrote:> yes "performance.parallel-readdir on and 1x3 replica >That's surprising. I thought performance.parallel-readdir will help only when distribute count is fairly high. This is something worth investigating further.> On Thu, Aug 30, 2018 at 5:00 PM Raghavendra Gowdappa <rgowdapp at redhat.com> > wrote: > >> >> >> On Thu, Aug 30, 2018 at 8:08 PM, Davide Obbi <davide.obbi at booking.com> >> wrote: >> >>> Thanks Amar, >>> >>> i have enabled the negative lookups cache on the volume: >>> >>I think enabling nl-cache-positive-entry might help for untarring or git clone into glusterfs. Its disabled by default. can you let us know the results? Option: performance.nl-cache-positive-entry Default Value: (null) Description: enable/disable storing of entries that were lookedup and found to be present in the volume, thus lookup on non existent file is served from the cache>>> To deflate a tar archive (not compressed) of 1.3GB it takes aprox 9mins >>> which can be considered a slight improvement from the previous 12-15 >>> however still not fast enough compared to local disk. The tar is present on >>> the gluster share/volume and deflated inside the same folder structure. >>> >> >> I am assuming this is with parallel-readdir enabled, right? >> >> >>> Running the operation twice (without removing the already deflated >>> files) also did not reduce the time spent. >>> >>> Running the operation with the tar archive on local disk made no >>> difference >>> >>> What really made a huge difference while git cloning was setting >>> "performance.parallel-readdir on". During the phase "Receiving objects" , >>> as i enabled the xlator it bumped up from 3/4MBs to 27MBs >>> >> >> What is the distribute count? Is it 1x3 replica? >> >> >>> So in conclusion i'm trying to make the untar operation working at an >>> acceptable level, not expecting local disks speed but at least being within >>> the 4mins >>> >>> I have attached the profiles collected at the end of the untar >>> operations with the archive on the mount and outside >>> >>> thanks >>> Davide >>> >>> >>> On Tue, Aug 28, 2018 at 8:41 AM Amar Tumballi <atumball at redhat.com> >>> wrote: >>> >>>> One of the observation we had with git clone like work load was, >>>> nl-cache (negative-lookup cache), helps here. >>>> >>>> Try 'gluster volume set $volume-name nl-cache enable'. >>>> >>>> Also sharing the 'profile info' during this performance observations >>>> also helps us to narrow down the situation. >>>> >>>> More on how to capture profile info @ https://hackmd.io/ >>>> PhhT5jPdQIKxzfeLQmnjJQ?view >>>> >>>> -Amar >>>> >>>> >>>> On Thu, Aug 23, 2018 at 7:11 PM, Davide Obbi <davide.obbi at booking.com> >>>> wrote: >>>> >>>>> Hello, >>>>> >>>>> did anyone ever managed to achieve reasonable waiting time while >>>>> performing metadata intensive operations such as git clone, untar etc...? >>>>> Is this possible workload or will never be in scope for glusterfs? >>>>> >>>>> I'd like to know, if possible, what would be the options that affect >>>>> such volume performances. >>>>> Albeit i managed to achieve decent git status/git grep operations, 3 >>>>> and 30 secs, the git clone and untarring a file from/to the same share take >>>>> ages. for a git repo of aprox 6GB. >>>>> >>>>> I'm running a test environment with 3 way replica 128GB RAM and 24 >>>>> cores are 2.40GHz, one internal SSD dedicated to the volume brick and 10Gb >>>>> network >>>>> >>>>> The options set so far that affects volume performances are: >>>>> 48 performance.readdir-ahead: on >>>>> 49 features.cache-invalidation-timeout: 600 >>>>> 50 features.cache-invalidation: on >>>>> 51 performance.md-cache-timeout: 600 >>>>> 52 performance.stat-prefetch: on >>>>> 53 performance.cache-invalidation: on >>>>> 54 performance.parallel-readdir: on >>>>> 55 network.inode-lru-limit: 900000 >>>>> 56 performance.io-thread-count: 32 >>>>> 57 performance.cache-size: 10GB >>>>> >>>>> _______________________________________________ >>>>> Gluster-users mailing list >>>>> Gluster-users at gluster.org >>>>> https://lists.gluster.org/mailman/listinfo/gluster-users >>>>> >>>> >>>> >>>> >>>> -- >>>> Amar Tumballi (amarts) >>>> >>> >>> >>> -- >>> Davide Obbi >>> System Administrator >>> >>> Booking.com B.V. >>> Vijzelstraat 66 >>> <https://maps.google.com/?q=Vijzelstraat+66&entry=gmail&source=g>-80 >>> Amsterdam 1017HL Netherlands >>> Direct +31207031558 >>> [image: Booking.com] <https://www.booking.com/> >>> The world's #1 accommodation site >>> 43 languages, 198+ offices worldwide, 120,000+ global destinations, >>> 1,550,000+ room nights booked every day >>> No booking fees, best price always guaranteed >>> Subsidiary of Booking Holdings Inc. (NASDAQ: BKNG) >>> >>> _______________________________________________ >>> Gluster-users mailing list >>> Gluster-users at gluster.org >>> https://lists.gluster.org/mailman/listinfo/gluster-users >>> >> >> > > -- > Davide Obbi > System Administrator > > Booking.com B.V. > Vijzelstraat 66 > <https://maps.google.com/?q=Vijzelstraat+66&entry=gmail&source=g>-80 > Amsterdam 1017HL Netherlands > Direct +31207031558 > [image: Booking.com] <https://www.booking.com/> > The world's #1 accommodation site > 43 languages, 198+ offices worldwide, 120,000+ global destinations, > 1,550,000+ room nights booked every day > No booking fees, best price always guaranteed > Subsidiary of Booking Holdings Inc. (NASDAQ: BKNG) >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.gluster.org/pipermail/gluster-users/attachments/20180830/81406c7c/attachment.html>
Poornima Gurusiddaiah
2018-Aug-30 15:47 UTC
[Gluster-users] [External] Re: file metadata operations performance - gluster 4.1
To enable nl-cache please use group option instead of single volume set: #gluster vol set VOLNAME group nl-cache This sets few other things including time out, invalidation etc. For enabling the option Raghavendra mentioned, you ll have to execute it explicitly, as it's not part of group option yet: #gluster vol set VOLNAME performance.nl-cache-positive-entry on Also from the past experience, setting the below option has helped in performance: # gluster vol set VOLNAME network.inode-lru-limit 200000 Regards, Poornima On Thu, Aug 30, 2018, 8:49 PM Raghavendra Gowdappa <rgowdapp at redhat.com> wrote:> > > On Thu, Aug 30, 2018 at 8:38 PM, Davide Obbi <davide.obbi at booking.com> > wrote: > >> yes "performance.parallel-readdir on and 1x3 replica >> > > That's surprising. I thought performance.parallel-readdir will help only > when distribute count is fairly high. This is something worth investigating > further. > > >> On Thu, Aug 30, 2018 at 5:00 PM Raghavendra Gowdappa <rgowdapp at redhat.com> >> wrote: >> >>> >>> >>> On Thu, Aug 30, 2018 at 8:08 PM, Davide Obbi <davide.obbi at booking.com> >>> wrote: >>> >>>> Thanks Amar, >>>> >>>> i have enabled the negative lookups cache on the volume: >>>> >>> > I think enabling nl-cache-positive-entry might help for untarring or git > clone into glusterfs. Its disabled by default. can you let us know the > results? > > Option: performance.nl-cache-positive-entry > Default Value: (null) > Description: enable/disable storing of entries that were lookedup and > found to be present in the volume, thus lookup on non existent file is > served from the cache > > >>>> To deflate a tar archive (not compressed) of 1.3GB it takes aprox 9mins >>>> which can be considered a slight improvement from the previous 12-15 >>>> however still not fast enough compared to local disk. The tar is present on >>>> the gluster share/volume and deflated inside the same folder structure. >>>> >>> >>> I am assuming this is with parallel-readdir enabled, right? >>> >>> >>>> Running the operation twice (without removing the already deflated >>>> files) also did not reduce the time spent. >>>> >>>> Running the operation with the tar archive on local disk made no >>>> difference >>>> >>>> What really made a huge difference while git cloning was setting >>>> "performance.parallel-readdir on". During the phase "Receiving objects" , >>>> as i enabled the xlator it bumped up from 3/4MBs to 27MBs >>>> >>> >>> What is the distribute count? Is it 1x3 replica? >>> >>> >>>> So in conclusion i'm trying to make the untar operation working at an >>>> acceptable level, not expecting local disks speed but at least being within >>>> the 4mins >>>> >>>> I have attached the profiles collected at the end of the untar >>>> operations with the archive on the mount and outside >>>> >>>> thanks >>>> Davide >>>> >>>> >>>> On Tue, Aug 28, 2018 at 8:41 AM Amar Tumballi <atumball at redhat.com> >>>> wrote: >>>> >>>>> One of the observation we had with git clone like work load was, >>>>> nl-cache (negative-lookup cache), helps here. >>>>> >>>>> Try 'gluster volume set $volume-name nl-cache enable'. >>>>> >>>>> Also sharing the 'profile info' during this performance observations >>>>> also helps us to narrow down the situation. >>>>> >>>>> More on how to capture profile info @ >>>>> https://hackmd.io/PhhT5jPdQIKxzfeLQmnjJQ?view >>>>> >>>>> -Amar >>>>> >>>>> >>>>> On Thu, Aug 23, 2018 at 7:11 PM, Davide Obbi <davide.obbi at booking.com> >>>>> wrote: >>>>> >>>>>> Hello, >>>>>> >>>>>> did anyone ever managed to achieve reasonable waiting time while >>>>>> performing metadata intensive operations such as git clone, untar etc...? >>>>>> Is this possible workload or will never be in scope for glusterfs? >>>>>> >>>>>> I'd like to know, if possible, what would be the options that affect >>>>>> such volume performances. >>>>>> Albeit i managed to achieve decent git status/git grep operations, 3 >>>>>> and 30 secs, the git clone and untarring a file from/to the same share take >>>>>> ages. for a git repo of aprox 6GB. >>>>>> >>>>>> I'm running a test environment with 3 way replica 128GB RAM and 24 >>>>>> cores are 2.40GHz, one internal SSD dedicated to the volume brick and 10Gb >>>>>> network >>>>>> >>>>>> The options set so far that affects volume performances are: >>>>>> 48 performance.readdir-ahead: on >>>>>> 49 features.cache-invalidation-timeout: 600 >>>>>> 50 features.cache-invalidation: on >>>>>> 51 performance.md-cache-timeout: 600 >>>>>> 52 performance.stat-prefetch: on >>>>>> 53 performance.cache-invalidation: on >>>>>> 54 performance.parallel-readdir: on >>>>>> 55 network.inode-lru-limit: 900000 >>>>>> 56 performance.io-thread-count: 32 >>>>>> 57 performance.cache-size: 10GB >>>>>> >>>>>> _______________________________________________ >>>>>> Gluster-users mailing list >>>>>> Gluster-users at gluster.org >>>>>> https://lists.gluster.org/mailman/listinfo/gluster-users >>>>>> >>>>> >>>>> >>>>> >>>>> -- >>>>> Amar Tumballi (amarts) >>>>> >>>> >>>> >>>> -- >>>> Davide Obbi >>>> System Administrator >>>> >>>> Booking.com B.V. >>>> Vijzelstraat 66 >>>> <https://maps.google.com/?q=Vijzelstraat+66&entry=gmail&source=g>-80 >>>> Amsterdam 1017HL Netherlands >>>> Direct +31207031558 >>>> [image: Booking.com] <https://www.booking.com/> >>>> The world's #1 accommodation site >>>> 43 languages, 198+ offices worldwide, 120,000+ global destinations, >>>> 1,550,000+ room nights booked every day >>>> No booking fees, best price always guaranteed >>>> Subsidiary of Booking Holdings Inc. (NASDAQ: BKNG) >>>> >>>> _______________________________________________ >>>> Gluster-users mailing list >>>> Gluster-users at gluster.org >>>> https://lists.gluster.org/mailman/listinfo/gluster-users >>>> >>> >>> >> >> -- >> Davide Obbi >> System Administrator >> >> Booking.com B.V. >> Vijzelstraat 66 >> <https://maps.google.com/?q=Vijzelstraat+66&entry=gmail&source=g>-80 >> Amsterdam 1017HL Netherlands >> Direct +31207031558 >> [image: Booking.com] <https://www.booking.com/> >> The world's #1 accommodation site >> 43 languages, 198+ offices worldwide, 120,000+ global destinations, >> 1,550,000+ room nights booked every day >> No booking fees, best price always guaranteed >> Subsidiary of Booking Holdings Inc. (NASDAQ: BKNG) >> > > _______________________________________________ > Gluster-users mailing list > Gluster-users at gluster.org > https://lists.gluster.org/mailman/listinfo/gluster-users-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.gluster.org/pipermail/gluster-users/attachments/20180830/c7c9efbd/attachment.html>