Re: Latency analysis of GlusterFS' network layer for pgbench

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On Sat, Jan 26, 2019 at 8:03 AM Raghavendra Gowdappa <rgowdapp@xxxxxxxxxx> wrote:


On Fri, Jan 11, 2019 at 8:09 PM Raghavendra Gowdappa <rgowdapp@xxxxxxxxxx> wrote:
Here is the update of the progress till now:
* The client profile attached till now shows the tuple creation is dominated by writes and fstats. Note that fstats are side-effects of writes as writes invalidate attributes of the file from kernel attribute cache.
* The rest of the init phase (which is marked by msgs "setting primary key" and "vaccuum") is dominated by reads. Next bigger set of operations are writes followed by fstats.

So, only writes, reads and fstats are the operations we need to optimize to reduce the init time latency. As mentioned in my previous mail, I did following tunings:
* Enabled only write-behind, md-cache and open-behind.
    - write-behind was configured with a cache-size/window-size of 20MB
    - open-behind was configured with read-after-open yes
    - md-cache was loaded as a child of write-behind in xlator graph. As a parent of write-behind, writes responses of writes cached in write-behind would invalidate stats. But when loaded as a child of write-behind this problem won't be there. Note that in both cases fstat would pass through write-behind (In the former case due to no stats in md-cache). However in the latter case fstats can be served by md-cache.
    - md-cache used to aggressively invalidate inodes. For the purpose of this test, I just commented out inode-invalidate code in md-cache. We need to fine tune the invalidation invocation logic.
    - set group-metadata-cache to on. But turned off upcall notifications. Note that since this workload basically accesses all its data through single mount point. So, there is no shared files across mounts and hence its safe to turn off invalidations.
With the above set of tunings I could reduce the init time of scale 8000 from 16.6 hrs to 11.4 hrs - an improvement in the range 25% to 30%

Since the workload is dominated by reads, we think a good read-cache where reads to regions just written are served from cache would greatly improve the performance. Since kernel page-cache already provides that functionality along with read-ahead (which is more intelligent and serves more read patterns than supported by Glusterfs read-ahead), we wanted to try that. But, Manoj found a bug where reads followed by writes are not served from page cache [5]. I am currently waiting for the resolution of this bug. As an alternative, I can modify io-cache to serve reads from the data just written. But, the change involves its challenges and hence would like to get a resolution on [5] (either positive or negative) before proceeding with modifications to io-cache.

As to the rpc latency, Krutika had long back identified that reading a single rpc message involves atleast 4 reads to socket. These many number of reads were done to identify the structure of the message on the go. The reason we wanted to discover the rpc message was to identify the part of the rpc message containing read or write payload and make sure that payload is directly read into a buffer different than the one containing rest of the rpc message. This strategy will make sure payloads are not copied again when buffers are moved across caches (read-ahead, io-cache etc) and also the rest of the rpc message can be freed even though the payload outlives the rpc message (when payloads are cached). However, we can experiment an approach where we can either do away with zero-copy requirement or let the entire buffer containing rpc message and payload to live in the cache.

From my observations and discussions with Manoj and Xavi, this workload is very sensitive to latency (than to concurrency). So, I am hopeful the above approaches will give positive results.

Me, Manoj and Csaba figured out that invalidations by md-cache and Fuse auto-invalidations  were dropping the kernel page-cache (more details on [5]). Changes to stats by writes from same client (local writes) were triggering both these codepaths dropping the cache. Since all the I/O done by this workload goes through the caches of single client, the invalidations are not necessary and I made code changes to fuse-bridge to disable auto-invalidations completely and commented out inode-invalidations in md-cache. Note that this doesn't regress the consistency/coherency of data seen in the caches as its a single client use-case. With these two changes coupled with earlier optimizations (client-io-threads=on, server/client-event-threads=4, md-cache as a child of write-behind in xlator graph, performance.md-cache-timeout=600), pgbench init of scale 8000 on a volume with NVMe backend completed in 54m25s. This is a whopping 94% improvement to the time we started out with (59280s vs 3360s).

These numbers were taken from the latest run I had scheduled. However, I didn't notice that the test had failed midway. From another test that had completed successfully, the numbers are 139m7s. That will be an improvement of 86% (59280s vs 8340s). I've scheduled another run just to be sure. The improvement is 86% and not 94%.


[root@shakthi4 ~]# gluster volume info
 
Volume Name: nvme-r3
Type: Replicate
Volume ID: d1490bcc-bcf1-4e09-91e8-ab01d9781263
Status: Started
Snapshot Count: 0
Number of Bricks: 1 x 3 = 3
Transport-type: tcp
Bricks:
Brick1: shakthi4:/gluster/nvme0n1/bricks/nvme-r3-1
Brick2: shakthi4:/gluster/nvme0n1/bricks/nvme-r3-2
Brick3: shakthi4:/gluster/nvme0n1/bricks/nvme-r3-3
Options Reconfigured:
server.event-threads: 4
client.event-threads: 4
diagnostics.client-log-level: INFO
performance.md-cache-timeout: 600
performance.io-cache: off
performance.read-ahead: off
diagnostics.count-fop-hits: on
diagnostics.latency-measurement: on
transport.address-family: inet
nfs.disable: on
performance.client-io-threads: on
performance.stat-prefetch: on

I'll be concentrating on how to disable fuse-auto-invalidations without regressing on the consistency model we've been providing till now. The consistency model Glusterfs has been providing till now is close to open consistency similar to what NFS provides [6][7].

But the initial thoughts are, at least for the pgbench test-case there is no harm in totally disabling fuse-auto-invalidations and md-cache invalidations as this workload totally runs on single mount point and hence invalidations itself are not necessary as all I/O goes through caches and hence caches are in sync with the state of the file on backend.

[6] http://nfs.sourceforge.net/#faq_a8
[7] https://lists.gluster.org/pipermail/gluster-users/2013-March/012805.html


On Fri, Dec 28, 2018 at 12:44 PM Raghavendra Gowdappa <rgowdapp@xxxxxxxxxx> wrote:


On Mon, Dec 24, 2018 at 6:05 PM Raghavendra Gowdappa <rgowdapp@xxxxxxxxxx> wrote:


On Mon, Dec 24, 2018 at 3:40 PM Sankarshan Mukhopadhyay <sankarshan.mukhopadhyay@xxxxxxxxx> wrote:
[pulling the conclusions up to enable better in-line]

> Conclusions:
>
> We should never have a volume with caching-related xlators disabled. The price we pay for it is too high. We need to make them work consistently and aggressively to avoid as many requests as we can.

Are there current issues in terms of behavior which are known/observed
when these are enabled?

We did have issues with pgbench in past. But they've have been fixed. Please refer to bz [1] for details. On 5.1, it runs successfully with all caching related xlators enabled. Having said that the only performance xlators which gave improved performance were open-behind and write-behind [2] (write-behind had some issues, which will be fixed by [3] and we'll have to measure performance again with fix to [3]).

One quick update. Enabling write-behind and md-cache with fix for [3] reduced the total time taken for pgbench init phase roughly by 20%-25% (from 12.5 min to 9.75 min for a scale of 100). Though this is still a huge time (around 12hrs for a db of scale 8000). I'll follow up with a detailed report once my experiments are complete. Currently trying to optimize the read path.
 
For some reason, read-side caching didn't improve transactions per second. I am working on this problem currently. Note that these bugs measure transaction phase of pgbench, but what xavi measured in his mail is init phase. Nevertheless, evaluation of read caching (metadata/data) will still be relevant for init phase too.

[1] https://bugzilla.redhat.com/show_bug.cgi?id=1512691
[2] https://bugzilla.redhat.com/show_bug.cgi?id=1629589#c4
[3] https://bugzilla.redhat.com/show_bug.cgi?id=1648781


> We need to analyze client/server xlators deeper to see if we can avoid some delays. However optimizing something that is already at the microsecond level can be very hard.

That is true - are there any significant gains which can be accrued by
putting efforts here or, should this be a lower priority?

The problem identified by xavi is also the one we (Manoj, Krutika, me and Milind) had encountered in the past [4]. The solution we used was to have multiple rpc connections between single brick and client. The solution indeed fixed the bottleneck. So, there is definitely work involved here - either to fix the single connection model or go with multiple connection model. Its preferred to improve single connection and resort to multiple connections only if bottlenecks in single connection are not fixable. Personally I think this is high priority along with having appropriate client side caching.

[4] https://bugzilla.redhat.com/show_bug.cgi?id=1467614#c52


> We need to determine what causes the fluctuations in brick side and avoid them.
> This scenario is very similar to a smallfile/metadata workload, so this is probably one important cause of its bad performance.

What kind of instrumentation is required to enable the determination?

On Fri, Dec 21, 2018 at 1:48 PM Xavi Hernandez <xhernandez@xxxxxxxxxx> wrote:
>
> Hi,
>
> I've done some tracing of the latency that network layer introduces in gluster. I've made the analysis as part of the pgbench performance issue (in particulat the initialization and scaling phase), so I decided to look at READV for this particular workload, but I think the results can be extrapolated to other operations that also have small latency (cached data from FS for example).
>
> Note that measuring latencies introduces some latency. It consists in a call to clock_get_time() for each probe point, so the real latency will be a bit lower, but still proportional to these numbers.
>

[snip]
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-devel

[Index of Archives]     [Gluster Users]     [Ceph Users]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux