I too struggle with speed issues in hci. Latency is a big problem with writes for me especially when dealing with small file workloads. How are you testing exactly? Look into enabling libgfapi and try some comparisons with that. People have been saying it’s much faster, but it’s not a default option and has a few bugs. Redhat devs do not appear to be giving its implementation any priority unfortunately. I’ve been considering switching to nfs storage because I’m seeing much better performance in testing with it. I have some nvme drives on the way and am curious how they would perform in hci but I’m thinking the issue is not a disk bottleneck (that appears very obvious in your case as well) On Mon, Mar 23, 2020 at 6:44 PM Christian Reiss <email(a)christian-reiss.de&gt; wrote: > Hey folks, > > gluster related question. Having SSD in a RAID that can do 2 GB writes > and Reads (actually above, but meh) in a 3-way HCI cluster connected > with 10gbit connection things are pretty slow inside gluster. > > I have these settings: > > Options Reconfigured: > cluster.server-quorum-type: server > cluster.quorum-type: auto > cluster.shd-max-threads: 8 > features.shard: on > features.shard-block-size: 64MB > server.event-threads: 8 > user.cifs: off > cluster.shd-wait-qlength: 10000 > cluster.locking-scheme: granular > cluster.eager-lock: enable > performance.low-prio-threads: 32 > network.ping-timeout: 30 > cluster.granular-entry-heal: enable > storage.owner-gid: 36 > storage.owner-uid: 36 > cluster.choose-local: true > client.event-threads: 16 > performance.strict-o-direct: on > network.remote-dio: enable > performance.client-io-threads: on > nfs.disable: on > storage.fips-mode-rchecksum: on > transport.address-family: inet > cluster.readdir-optimize: on > cluster.metadata-self-heal: on > cluster.data-self-heal: on > cluster.entry-self-heal: on > cluster.data-self-heal-algorithm: full > features.uss: enable > features.show-snapshot-directory: on > features.barrier: disable > auto-delete: enable > snap-activate-on-create: enable > > Writing inside the /gluster_bricks yields those 2GB/sec writes, Reading > the same. > > Reading inside the /rhev/data-center/mnt/glusterSD/ dir reads go down to > 366mb/sec while writes plummet to to 200mb/sec. > > Summed up: Writing into the SSD Raid in the lvm/xfs gluster brick > directory is fast, writing into the mounted gluster dir is horribly slow. > > The above can be seen and repeated on all 3 servers. The network can do > full 10gbit (tested with, among others: rsync, iperf3). > > Anyone with some idea on whats missing/ going on here? > > Thanks folks, > as always stay safe and healthy! > > -- > with kind regards, > mit freundlichen Gruessen, > > Christian Reiss > _______________________________________________ > Users mailing list -- users(a)ovirt.org > To unsubscribe send an email to users-leave(a)ovirt.org > Privacy Statement: https://www.ovirt.org/privacy-policy.html > oVirt Code of Conduct: > https://www.ovirt.org/community/about/community-guidelines/ > List Archives: > https://lists.ovirt.org/archives/list/users@ovirt.org/message/OMAAERV4IUI... >

Hey Strahil, seems you're the go-to-guy with pretty much all my issues. I thank you for this and your continued support. Much appreciated. 200mb/reads however seems like a broken config or malfunctioning gluster than requiring performance tweaks. I enabled profiling so I have real life data available. But seriously even without tweaks I would like (need) 4 times those numbers, 800mb write speed is okay'ish, given the fact that 10gbit backbone can be the limiting factor. We are running BigCouch/CouchDB Applications that really really need IO. Not in throughput but in response times. 200mb/s is just way off. It feels as gluster can/should do more, natively. -Chris. On 24/03/2020 06:17, Strahil Nikolov wrote: > Hey Chris,, > > You got some options. > 1. To speedup the reads in HCI - you can use the option : > cluster.choose-local: on > 2. You can adjust the server and client event-threads > 3. You can use NFS Ganesha (which connects to all servers via libgfapi) as a NFS Server. > In such case you have to use some clustering like ctdb or pacemaker. > Note:disable cluster.choose-local if you use this one > 4 You can try the built-in NFS , although it's deprecated (NFS Ganesha is fully supported) > 5. Create a gluster profile during the tests. I have seen numerous improperly selected tests -> so test with real-world workload. Synthetic tests are not good. > > Best Regards, > Strahil Nikolov

Christian, Adding on to Stahil’s notes, make sure you’re using jumbo MTUs on servers and client host nodes. Making sure you’re using appropriate disk schedulers on hosts and VMs is important, worth double checking that it’s doing what you think it is. If you are only HCI, gluster’s choose-local on is a good thing, but try cluster.choose-local: false cluster.read-hash-mode: 3 if you have separate servers or nodes with are not HCI to allow it spread reads over multiple nodes. Test out these settings if you have lots of RAM and cores on your servers, they work well for me with 20 cores and 64GB ram on my servers with my load: performance.io-thread-count: 64 performance.low-prio-threads: 32 these are worth testing for your workload. If you’re running VMs with these, test out libglapi connections, it’s significantly better for IO latency than plain fuse mounts. If you can tolerate the issues, the biggest one at the moment being you can’t take snapshots of the VMs with it enabled as of March. If you have tuned available, I use throughput-performance on my servers and guest-host on my vm nodes, throughput-performance on some HCI ones. I’d test with out the fips-rchecksum setting, that may be creating extra work for your servers. If you mounted individual bricks, check that you disabled barriers on them at mount if appropriate. Hope it helps, -Darrell > On Mar 24, 2020, at 6:23 AM, Strahil Nikolov <hunter86_bg(a)yahoo.com&gt; wrote: > > On March 24, 2020 11:20:10 AM GMT+02:00, Christian Reiss <email(a)christian-reiss.de&gt; wrote: >> Hey Strahil, >> >> seems you're the go-to-guy with pretty much all my issues. I thank you >> for this and your continued support. Much appreciated. >> >> >> 200mb/reads however seems like a broken config or malfunctioning >> gluster >> than requiring performance tweaks. I enabled profiling so I have real >> life data available. But seriously even without tweaks I would like >> (need) 4 times those numbers, 800mb write speed is okay'ish, given the >> fact that 10gbit backbone can be the limiting factor. >> >> We are running BigCouch/CouchDB Applications that really really need >> IO. >> Not in throughput but in response times. 200mb/s is just way off. >> >> It feels as gluster can/should do more, natively. >> >> -Chris. >> >> On 24/03/2020 06:17, Strahil Nikolov wrote: >>> Hey Chris,, >>> >>> You got some options. >>> 1. To speedup the reads in HCI - you can use the option : >>> cluster.choose-local: on >>> 2. You can adjust the server and client event-threads >>> 3. You can use NFS Ganesha (which connects to all servers via >> libgfapi) as a NFS Server. >>> In such case you have to use some clustering like ctdb or pacemaker. >>> Note:disable cluster.choose-local if you use this one >>> 4 You can try the built-in NFS , although it's deprecated (NFS >> Ganesha is fully supported) >>> 5. Create a gluster profile during the tests. I have seen numerous >> improperly selected tests -> so test with real-world workload. >> Synthetic tests are not good. >>> >>> Best Regards, >>> Strahil Nikolov > > Hey Chris, > > What type is your VM ? > Try with 'High Performance' one (there is a good RH documentation on that topic). > > If the DB load was directly on gluster, you could use the settings in the '/var/lib/gluster/groups/db-workload' to optimize that, but I'm not sure if this will bring any performance on a VM. > > 1. Check the VM disk scheduler. Use 'noop/none' (depends on multiqueue is enabled) to allow the Hypervisor aggregate the I/O requests from multiple VMs. > Next, set 'noop/none' disk scheduler on the hosts - these 2 are the optimal for SSDs and NVME disks (if I recall corectly you are using SSDs) > > 2. Disable cstates on the host and Guest (there are a lot of articles about that) > > 3. Enable MTU 9000 for Hypervisor (gluster node). > > 4. You can try setting/unsetting the tunables in the db-workload group and run benchmarks with real workload . > > 5. Some users reported that enabling TCP offload on the hosts gave huge improvement in performance of gluster - you can try that. > Of course there are mixed feelings - as others report that disabling it brings performance. I guess it is workload specific. > > 6. You can try to tune the 'performance.readahead' on your gluster volume. > > Here are some settings of some users /from an old e-mail/: > > performance.read-ahead: on > performance.stat-prefetch: on > performance.flush-behind: on > performance.client-io-threads: on > performance.write-behind-window-size: 64MB (shard size) > > > > For a 48 cores / host: > > server.event-threads: 4 > client.event-threads: 8 > > Your ecent-threads seem to be too high.And yes, documentation explains it , but without an example it becomes more confusing. > > Best Regards, > Strahil Nikolov > > > _______________________________________________ > Users mailing list -- users(a)ovirt.org > To unsubscribe send an email to users-leave(a)ovirt.org > Privacy Statement: https://www.ovirt.org/privacy-policy.html > oVirt Code of Conduct: https://www.ovirt.org/community/about/community-guidelines/ > List Archives: https://lists.ovirt.org/archives/list/users@ovirt.org/message/BOFZEJPBIRX...

Red hat also recommends a shard size of 512mb, it's actually the only shard size they support. Also check the chunk size on the LVM thin pools running the bricks, should be at least 2mb. Note that changing the shard size only applies to new VM disks after the change. Changing the chunk size requires making a new brick. libgfapi brings a huge performance boost, in my opinion its almost a necessity unless you have a ton of extra disk speed / network throughput. Just be aware of the caveats. On 2020-03-24 14:12, Strahil Nikolov wrote: > On March 24, 2020 7:33:16 PM GMT+02:00, Darrell Budic > <budic(a)onholyground.com&gt; wrote: >> Christian, >> >> Adding on to Stahil’s notes, make sure you’re using jumbo MTUs on >> servers and client host nodes. Making sure you’re using appropriate >> disk schedulers on hosts and VMs is important, worth double checking >> that it’s doing what you think it is. If you are only HCI, gluster’s >> choose-local on is a good thing, but try >> >> cluster.choose-local: false >> cluster.read-hash-mode: 3 >> >> if you have separate servers or nodes with are not HCI to allow it >> spread reads over multiple nodes. >> >> Test out these settings if you have lots of RAM and cores on your >> servers, they work well for me with 20 cores and 64GB ram on my >> servers >> with my load: >> >> performance.io-thread-count: 64 >> performance.low-prio-threads: 32 >> >> these are worth testing for your workload. >> >> If you’re running VMs with these, test out libglapi connections, it’s >> significantly better for IO latency than plain fuse mounts. If you can >> tolerate the issues, the biggest one at the moment being you can’t >> take >> snapshots of the VMs with it enabled as of March. >> >> If you have tuned available, I use throughput-performance on my >> servers >> and guest-host on my vm nodes, throughput-performance on some HCI >> ones. >> >> >> I’d test with out the fips-rchecksum setting, that may be creating >> extra work for your servers. >> >> If you mounted individual bricks, check that you disabled barriers on >> them at mount if appropriate. >> >> Hope it helps, >> >> -Darrell >> >>> On Mar 24, 2020, at 6:23 AM, Strahil Nikolov <hunter86_bg(a)yahoo.com&gt; >> wrote: >>> >>> On March 24, 2020 11:20:10 AM GMT+02:00, Christian Reiss >> <email(a)christian-reiss.de&gt; wrote: >>>> Hey Strahil, >>>> >>>> seems you're the go-to-guy with pretty much all my issues. I thank >> you >>>> for this and your continued support. Much appreciated. >>>> >>>> >>>> 200mb/reads however seems like a broken config or malfunctioning >>>> gluster >>>> than requiring performance tweaks. I enabled profiling so I have >> real >>>> life data available. But seriously even without tweaks I would like >>>> (need) 4 times those numbers, 800mb write speed is okay'ish, given >> the >>>> fact that 10gbit backbone can be the limiting factor. >>>> >>>> We are running BigCouch/CouchDB Applications that really really need >>>> IO. >>>> Not in throughput but in response times. 200mb/s is just way off. >>>> >>>> It feels as gluster can/should do more, natively. >>>> >>>> -Chris. >>>> >>>> On 24/03/2020 06:17, Strahil Nikolov wrote: >>>>> Hey Chris,, >>>>> >>>>> You got some options. >>>>> 1. To speedup the reads in HCI - you can use the option : >>>>> cluster.choose-local: on >>>>> 2. You can adjust the server and client event-threads >>>>> 3. You can use NFS Ganesha (which connects to all servers via >>>> libgfapi) as a NFS Server. >>>>> In such case you have to use some clustering like ctdb or >> pacemaker. >>>>> Note:disable cluster.choose-local if you use this one >>>>> 4 You can try the built-in NFS , although it's deprecated (NFS >>>> Ganesha is fully supported) >>>>> 5. Create a gluster profile during the tests. I have seen numerous >>>> improperly selected tests -> so test with real-world workload. >>>> Synthetic tests are not good. >>>>> >>>>> Best Regards, >>>>> Strahil Nikolov >>> >>> Hey Chris, >>> >>> What type is your VM ? >>> Try with 'High Performance' one (there is a good RH documentation on >> that topic). >>> >>> If the DB load was directly on gluster, you could use the settings >> in the '/var/lib/gluster/groups/db-workload' to optimize that, but >> I'm >> not sure if this will bring any performance on a VM. >>> >>> 1. Check the VM disk scheduler. Use 'noop/none' (depends on >> multiqueue is enabled) to allow the Hypervisor aggregate the I/O >> requests from multiple VMs. >>> Next, set 'noop/none' disk scheduler on the hosts - these 2 are the >> optimal for SSDs and NVME disks (if I recall corectly you are using >> SSDs) >>> >>> 2. Disable cstates on the host and Guest (there are a lot of articles >> about that) >>> >>> 3. Enable MTU 9000 for Hypervisor (gluster node). >>> >>> 4. You can try setting/unsetting the tunables in the db-workload >> group and run benchmarks with real workload . >>> >>> 5. Some users reported that enabling TCP offload on the hosts >> gave huge improvement in performance of gluster - you can try that. >>> Of course there are mixed feelings - as others report that >> disabling it brings performance. I guess it is workload specific. >>> >>> 6. You can try to tune the 'performance.readahead' on your >> gluster volume. >>> >>> Here are some settings of some users /from an old e-mail/: >>> >>> performance.read-ahead: on >>> performance.stat-prefetch: on >>> performance.flush-behind: on >>> performance.client-io-threads: on >>> performance.write-behind-window-size: 64MB (shard size) >>> >>> >>> >>> For a 48 cores / host: >>> >>> server.event-threads: 4 >>> client.event-threads: 8 >>> >>> Your ecent-threads seem to be too high.And yes, documentation >> explains it , but without an example it becomes more confusing. >>> >>> Best Regards, >>> Strahil Nikolov >>> >>> >>> _______________________________________________ >>> Users mailing list -- users(a)ovirt.org >>> To unsubscribe send an email to users-leave(a)ovirt.org >>> Privacy Statement: https://www.ovirt.org/privacy-policy.html >>> oVirt Code of Conduct: >> https://www.ovirt.org/community/about/community-guidelines/ >>> List Archives: >> https://lists.ovirt.org/archives/list/users@ovirt.org/message/BOFZEJPBIRX... > > When talking about mounts, you can avoid SELINUX lookups via > 'context=system_u:object_r:glusterd_brick_t:s0' mount option for all > bricks. > This way the kernel will reduce the requests to the bricks. > > Also 'noatime' is a default mount option(relatime is also a good one) > for HCI gluster bricks. > > > It seems you have a lot of checks to do :) > > Best Regards, > Strahil Nikolov > _______________________________________________ > Users mailing list -- users(a)ovirt.org > To unsubscribe send an email to users-leave(a)ovirt.org > Privacy Statement: https://www.ovirt.org/privacy-policy.html > oVirt Code of Conduct: > https://www.ovirt.org/community/about/community-guidelines/ > List Archives: > https://lists.ovirt.org/archives/list/users@ovirt.org/message/B7OSBWD4MGB... _______________________________________________ Users mailing list -- users(a)ovirt.org To unsubscribe send an email to users-leave(a)ovirt.org Privacy Statement: https://www.ovirt.org/privacy-policy.html oVirt Code of Conduct: https://www.ovirt.org/community/about/community-guidelines/ List Archives: https://lists.ovirt.org/archives/list/users@ovirt.org/message/PDG34NP2ADF...

Hey Jayme, thanks for replying; sorry for the delay. If I am understanding this right, there is no real official way to enable libgfapi. If you somehow manage to get it running then you will lose HA capabilities, which is something we like on our production servers. The most recent post I could find on the matter (https://www.mail-archive.com/users@ovirt.org/msg59664.html) read like its worth a try for hobyyists, but for production servers I do am a little bit scared. Do you maybe have any document or other source that does work with 4.3.x versions and inspires confidence? :-) -Chris On 24/03/2020 19:49, Jayme wrote: > I strongly believe that FUSE mount is the real reason for poor > performance in HCI and these minor gluster and other tweaks won't > satisfy most seeking i/o performance. Enabling libgfapi is probably the > best option. Redhat has recently closed bug reports related to libgfapi > citing won't fix and one comment suggests that libgfapi was not showing > good enough performance to bother with which appears to contradict what > many oVirt users are seeing. It's confusing to me why libgfapi as a > default option is not being given any priority. > > https://bugzilla.redhat.com/show_bug.cgi?id=1465810 > > "We do not plan to enable libgfapi for oVirt/RHV. We did not find enough > performance improvement justification for it"

Hey Alex, you too, thanks for writing. I'm on 64mb as per default for ovirt. We tried no sharding, 128mb sharding, 64mb sharding (always with copying the disk). There was no increase or decrease in disk speed in any way. Besides losing HA capabilites, what other caveats? -Chris. On 24/03/2020 19:25, Alex McWhirter wrote: > Red hat also recommends a shard size of 512mb, it's actually the only > shard size they support. Also check the chunk size on the LVM thin > pools running the bricks, should be at least 2mb. Note that changing > the shard size only applies to new VM disks after the change. Changing > the chunk size requires making a new brick. > > libgfapi brings a huge performance boost, in my opinion its almost a > necessity unless you have a ton of extra disk speed / network > throughput. Just be aware of the caveats. -- Christian Reiss - email(a)christian-reiss.de /"\ ASCII Ribbon support(a)alpha-labs.net \ / Campaign X against HTML WEB alpha-labs.net / \ in eMails GPG Retrieval https://gpg.christian-reiss.de GPG ID ABCD43C5, 0x44E29126ABCD43C5 GPG fingerprint = 9549 F537 2596 86BA 733C A4ED 44E2 9126 ABCD 43C5 "It's better to reign in hell than to serve in heaven.", John Milton, Paradise lost. _______________________________________________ Users mailing list -- users(a)ovirt.org To unsubscribe send an email to users-leave(a)ovirt.org Privacy Statement: https://www.ovirt.org/privacy-policy.html oVirt Code of Conduct: https://www.ovirt.org/community/about/community-guidelines/ List Archives: https://lists.ovirt.org/archives/list/users@ovirt.org/message/KL6HLEIRQ6G...

On 3/24/20 7:25 PM, Alex McWhirter wrote: > Red hat also recommends a shard size of 512mb, it's actually the only > shard size they support. Also check the chunk size on the LVM thin > pools running the bricks, should be at least 2mb. Note that changing > the shard size only applies to new VM disks after the change. Changing > the chunk size requires making a new brick. > Regarding the chunk size, red hat tells me it depends on RAID or JBOD https://access.redhat.com/documentation/en-us/red_hat_gluster_storage/3.5... chunksize An important parameter to be specified while creating a thin pool is the chunk size,which is the unit of allocation. For good performance, the chunk size for the thin pool and the parameters of the underlying hardware RAID storage should be chosen so that they work well together. And regarding the shard size, you can fix that with storage live migration right? Use two volumes and domains and move them so they will adopt the new shard size... Am I correct that when you change the sharding on a running volume, it only applies for new disks? Or does it also apply to extensions to a current disk? Met vriendelijke groet, With kind regards, Jorick Astrego Netbulae Virtualization Experts ---------------- Tel: 053 20 30 270 info(a)netbulae.eu Staalsteden 4-3A KvK 08198180 Fax: 053 20 30 271 www.netbulae.eu 7547 TA Enschede BTW NL821234584B01 ----------------

Hey, thanks for writing. If I go for dont choose local my speed drops dramatically (halving). Speed between the hosts is okay (tested) but for some odd reason the mtu is at 1500 still. I was sure I set it to jumbo/9k. Oh well. Not during runtime. I can hear the gluster scream if the network dies for a second :) -Chris. On 24/03/2020 18:33, Darrell Budic wrote: > > cluster.choose-local: false > cluster.read-hash-mode: 3 > > if you have separate servers or nodes with are not HCI to allow it > spread reads over multiple nodes. -- Christian Reiss - email(a)christian-reiss.de /"\ ASCII Ribbon support(a)alpha-labs.net \ / Campaign X against HTML WEB alpha-labs.net / \ in eMails GPG Retrieval https://gpg.christian-reiss.de GPG ID ABCD43C5, 0x44E29126ABCD43C5 GPG fingerprint = 9549 F537 2596 86BA 733C A4ED 44E2 9126 ABCD 43C5 "It's better to reign in hell than to serve in heaven.", John Milton, Paradise lost. _______________________________________________ Users mailing list -- users(a)ovirt.org To unsubscribe send an email to users-leave(a)ovirt.org Privacy Statement: https://www.ovirt.org/privacy-policy.html oVirt Code of Conduct: https://www.ovirt.org/community/about/community-guidelines/ List Archives: https://lists.ovirt.org/archives/list/users@ovirt.org/message/QYS7RIHXYAY...

Hey Strahil, as always: thanks! On 24/03/2020 12:23, Strahil Nikolov wrote: > performance.write-behind-window-size: 64MB (shard  size) This one doubled my speed from 200mb to 400mb!! I think this is where the meat is at. -Chris.

These settings mean: > performance.strict-o-direct: on > network.remote-dio: enable That you are using direct I/O both on the client and server side.

> Writing inside the /gluster_bricks yields those 2GB/sec writes, Reading > the same. How did you test this?

Did you test reading from the storage on the server side using direct I/O? if not, you test accessing server buffer cache, which is pretty fast.

> Reading inside the /rhev/data-center/mnt/glusterSD/ dir reads go down to > 366mb/sec while writes plummet to to 200mb/sec. This use direct I/O.

Please share the commands/configuration files used to perform the tests. Adding storage folks that can help with analyzing this.