Major Description
During performance testing of a new Lustre file system, we discovered that read/write performance aren't where we would expect. As an example, the block level read performance for the system is just over 65GB/s. In scaling tests, we can only get to around 30 GB/s for reads. Writes are slightly better, but still in the 35GB/s range. At single node scale, we seem to cap out at a few GB/s.
After going through tunings and everything that we can find, we're slightly better, but still miles behind where performance should be. We've played with various ksocklnd parameters (nconnds, nscheds, tx/rx buffer size, etc), but really to not much change. Current tunings that may be relevant: credits 2560, peer credits 63, max_rpcs_in_flight 32.
Network configuration on the servers is 2x 100G ethernet bonded together (active/active) using kernel bonding (not ksocklnd bonding).
iperf between two nodes gets nearly line rate at ~98Gb/s and iperf from two nodes to a single node can push ~190Gb/s, consistent with what would be expected from the kernel bonding.
lnet selftest shows about ~2.5GB/s (20Gb/s) rates for node to node tests. I'm not sure if this is a bug in lnet selftest or a real reflection of the performance.
We found the following related tickets/mailing list discussions which seem to be very similar to what we're seeing, but with no resolutions:
http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/2019-August/016630.html
https://jira.whamcloud.com/browse/LU-11415
https://jira.whamcloud.com/browse/LU-12815 (maybe performance limiting, but I doubt it for what we're seeing)
Any help or suggestions would be awesome.
Thanks!
- Jeff
Attachments
Issue Links
Activity
As a side note, it may make sense for us to close this particular issue and open a new, tailored one for the issues we're seeing now, since the particular issue described in this ticket (slow lnet performance) has been resolved with the `conns_per_peer` patch.
Along those lines, since that patch resolved our network performance issue and we'd like to keep running 2.12 (LTS), could we lobby to get James' backport of it included in the next 2.12 point release so that we don't have to keep carrying that patch?
Update on where we are:
When we stood up the system we ran some benchmarks on the raw block storage, so we're confident that the block storage can provide ~7GB/s read per LUN, with ~65GB/s read across the 12 LUNs in aggregate. What we did not do, however, was run any benchmarks on ZFS after the zpools were created on top of the LUN. Since LNET was no longer our bottleneck, we figured it would make sense to verify the stack from the bottom up, starting with the zpools. We set the zpools to `canmount=on` and changed the mountpoints, then mounted them and ran fio on them. Performance is terrible.
Given that we have another file system running with the exact same tunings and general layout, we also checked that file system in the same manner to much the same results. Since we have past benchmarking results from that file system, we're fairly confident that at some point in the past ZFS was functioning correctly. With that knowledge (and after looking at various zfs github issues) we decided to roll back from zfs 0.8.5 to 0.7.13 to test the performance there. It seems that 0.7.13 is also providing the same results.
I think that there may be potential value in rolling back our kernel to match what it was when we initialized the other file system, as there might be some odd interaction occurring with the kernel version we're running, but I'm not sure.
Here's the results of our testing on a single LUN with ZFS. Keep in mind this LUN can do ~7GB/s at the block level.
- files | read | write
1 file - 396 MB/s | 4.2 GB/s
4 files - 751 MB/s | 4.7 GB/s
12 files - 1.6 GB/s | 4.7 GB/s
And here's the really simple fio we're running to get these numbers:
fio --rw=read --size 20G --bs=1M --name=something --ioengine=libaio --runtime=60s --numjobs=12
We're also noticing some issues where Lustre is eating into those numbers significantly when layered on top. We're going to hold off on debugging that at all until zfs is stable though, as it may just be due to the same zfs issues.
Here's a table of performance values using lnet_selftest at various conns_per_peer values. I assume that this changes with CPU clock speed and other factors, but it at least shows that scaling for the patch commit message.
conns_per_peer setting - speed
1 - 1.7GiB/s
2 - 3.3GiB/s
4 - 6.4GiB/s
8 - 11.5GiB/s
16 - 11.5GiB/s
We did some more troubleshooting on our end yesterday and are suspecting some serious zfs issues. Currently testing an older ZFS version, and will comment again later after some more testing.
nilesj could you share the performance results for different conns_per_peer values? It would be useful to include a table with this information in the commit message for the patch.
As for the patch Amir mentioned, that was speculation regarding high CPU usage in osc_page_gang_lookup(). I can't definitively say whether that patch will help improve performance or not. Getting the flamegraphs for this would be very useful, along with what the test workload/parameters are (I'd assume IOR, but the options used are critical).
I chatted with Andreas about the read performance and he mentioned this:
there is patch https://review.whamcloud.com/40347 "LU-9920 vvp: dirty pages with pagevec" that is on master, but not 2.12 yet
what the write workload is doing is very important also. If there are large numbers of threads/clients writing to the same file, or the file is very large and there are lots of cached pages, or lots of separate DLM locks, then there is more work for osc_page_gang_lookup() to do. collecting perf stats for the workload and filing an LU ticket is probably the best way to go.
I would suggest trying this patch to see if there are any performance improvement. Would also be good to attach the flamegraphs we captured for both IOR writes and reads to get more eyes on it.
Sort of. It used 12 of the 24 configured threads. I've since reduced this, but wanted to mention what I was seeing in testing.
I performed quite a few more tests today with the LU-12815 patch applied and various tunings, and have some good news. With the patch, we can see nearly line rate with lnet selftest (11.5-12.0GB/s, up from ~2.5GB/s). Current tunings:
options ksocklnd sock_timeout=100 credits=2560 peer_credits=63 conns_per_peer=8 nscheds=12
8 conns_per_peer seemed to give the best performance, and nscheds had to be increased because I noticed that the 6 default threads were all 100% pegged during an lnet selftest.
Unfortunately, this isn't reflecting in the single node IOR numbers. While we saw a ~5x increase in the lnet selftest numbers, we're only seeing a 2x increase in IOR numbers. IOR writes went from ~5GB/s to 9.8GB/s and reads went from ~1.3GB/s to 2.6GB/s on a file-per-OST test (12 OSTs, 6 OSSs). Really trying to understand the brutal read disparity; hoping you all have some thoughts. The writes seem to prove that we can push that bandwidth over the network at least, but is there something about the read path that's different from a networking perspective?
Jeff, when you say "it only uses half", do you mean there are half the number of threads as when you configure nscheds to? If so, that's how it's suppose to work. The idea is not to consume all the cores with lnd threads, to allow other processes to use the system as well.
I did a back port of the LU-12815 work for 2.12 and we have full use of our Ethernet network.
Amir,
Are you talking about the socknal_sd01_xx threads? If so, I the work did span all of them. I just swapped to using a patched server/client with LU-12815 included and it seems that when I had the default 6, they were all being used, but if I increase `nscheds` to 24 (just matching core count), it only uses half. Really interesting behavior.
Jeff, another data point: When you switched to MR with virtual interfaces, was the load distributed to all the socklnd worker threads?
The reason I'm interested in this, is because the way work is assigned to the different CPTs is by hashing the NID. The Hash function will get us to one of the CPTs and then we pick one of the threads in that pool. If we have a single NID, we'll always get hashed into the same CPT, therefore we will not be utilizing all the worker threads. This could be another factor in the performance issue you're seeing.
If you could confirm the socklnd worker thread usage, that'll be great.
thanks
Jeff, I definitely have some comments related to ZFS performance, but it should really go into a separate ticket. If I file that ticket, it will not be tracked correctly as a customer issue, so it is best if you do that.
As for including conns_per_peer into 2.12, that is a bit tricky in the short term since that patch depends on another one that is removing the socklnd-level TCP bonding feature. While the LNet Multi-Rail provides better functionality, use_tcp_bonding may be in use at customer sites and shouldn't be removed in an LTS release without any warning. A patch will go into the next 2.12.7 LTS and 2.14.0 releases to announce that this option is deprecated, which will allow sites to become aware of this change and move over to LNet Multi-Rail. I've asked in
LU-12815for an email to be sent out to lustre-discuss and lustre-devel asking if anyone is using this feature, and maybe it can be removed from 2.12.8 if there is no feedback on its usage.