[Lustre-discuss] EXTERNAL: Re: LNET Performance Issue

Thu Feb 23 11:37:52 PST 2012

Hi all,

I'd suggest to start from simple point to point tests. There's too
many variables involved in a 'dd'. Please:

- Do a native IB write test from A to B, of 1M transfers, which is the
  max payload per Lustre RPC. With native IB bandwidth test tool, I
  remember there used to be an ib_write_bw tool, or something similar.

- Do a LNet selftest brw write test, size=1M, concurrency=8 and 16, from
  A to B. Be sure to check your lnet_selftest kernel module for an old
  option called srpc_peer_credits; if it's there, set it to 0. After
  the test is done, please run a "lctl --net o2ib conn_list" (replace
  o2ib with whatever @o2ib network in use). This command shows the
  path MTU for the connection, which might be of interest.

And let us know the results (command line options, selftest script,
and their outputs). Thanks.

- Isaac

On Tue, Feb 21, 2012 at 01:26:44PM +0000, Barberi, Carl E wrote:
> I am using QDR IB and I am not using any IB extenders.  This is a self-contained network, with no access outside of the lab it is currently in.
> 
> Carl
> 
> From: Jeremy Filizetti [mailto:jeremy.filizetti at gmail.com]
> Sent: Monday, February 20, 2012 8:37 PM
> To: Kevin Van Maren
> Cc: Barberi, Carl E; lustre-discuss at lists.Lustre.org; isaac_huang at xyratex.com
> Subject: Re: [Lustre-discuss] EXTERNAL: Re: LNET Performance Issue
> 
> It does seem extreme for data center IB latency but it may not be in the data center.  The LNet write should take 2 RTT latencies, and 3 for reads so you could double/triple those times plus any overhead.
> 
> Carl can you clarify if you are using QDR IB and/or any campus or wide area IB extenders?
> 
> Jeremy
> On Mon, Feb 20, 2012 at 8:14 PM, Kevin Van Maren <KVanMaren at fusionio.com<mailto:KVanMaren at fusionio.com>> wrote:
> While it's possible the default credits (8 as I recall) is not enough for peak performance, it seems to me that something else is wrong:
> Each 1MB RPC should take ~300uS (based on MPI/IB xfer rates of 3.2+ GB/s), so that means there is another 400uS overhead per RPC that is not masked with 8 concurrent RPCs, in addition to the overhead masked when he increased concurrency.  This is crazy, with a 1uS network latency.
> 
> Unless the RPCs are are being broken into tiny chunks or something -- does lnet do single-page xfers and not use a rendezvous protocol for full-sized RPCs?  It definitely seems that something is broken when o2iblnd gets ~1/3 of the MPI BW, given that lnd was designed for high-speed xfers.
> 
> The max_rpcs_in_flight normally needs tweaking to improve disk concurrency, where a single client needs to drive a high queue depth. Still finding it hard to believe 8 1MB concurrent RPCs can't handle the network.
> 
> Kevin
> 
> 
> On Feb 20, 2012, at 5:44 PM, "Jeremy Filizetti" <jeremy.filizetti at gmail.com<mailto:jeremy.filizetti at gmail.com>> wrote:
> Am I reading your earlier post correctly that you have a single server acting as the MDS and OSS?  Have you changed your peer_credits and credits for ko2iblnd kernel module on the server and client?  You also mentioned changing osc.*.max_dirty_mb, you probably need to adjust osc.*.max_rpcs_in_flight as well.  Can you post your rpc stats "lctl get_param osc.*.rpc_stats"?  I would guess they are bunching up around 7-8 if your running with the default max_rpcs_in_flight=8.
> 
> Jeremy
> 
> On Mon, Feb 20, 2012 at 4:59 PM, Barberi, Carl E <carl.e.barberi at lmco.com<mailto:carl.e.barberi at lmco.com>> wrote:
> Thank you.   This did help.  With the concurrency set to 16, I was able to get a max write speed of 1138 MB/s.  Any ideas on how we can make that faster, though?  Ideally, we'd like to get to 1.5 GB/s.
> 
> Carl
> 
> From: Liang Zhen [mailto:liang at whamcloud.com<mailto:liang at whamcloud.com>]
> Sent: Thursday, February 16, 2012 1:45 AM
> To: Barberi, Carl E
> Cc: 'lustre-discuss at lists.Lustre.org<mailto:lustre-discuss at lists.Lustre.org>'
> Subject: EXTERNAL: Re: [Lustre-discuss] LNET Performance Issue
> 
> Hi, I assume you are using "size=1M" for brw test right? performance could increase if you set "concurrency" while adding brw test, i.e: --concurrency=16
> 
> Liang
> 
> On Feb 16, 2012, at 3:30 AM, Barberi, Carl E wrote:
> 
> We are having issues with LNET performance over Infiniband.  We have a configuration with a single MDT and six (6) OSTs.  The Lustre client I am using to test is configured to use 6 stripes (lfs setstripe -c  6 /mnt/lustre).  When I perform a test using the following command:
> 
>                 dd if=/dev/zero of=/mnt/lustre/test.dat bs=1M count=2000
> 
> I typically get a write rate of about 815 MB/s, and we never exceed 848 MB/s.  When I run obdfilter-survey, we easily get about 3-4GB/s write speed, but when I run a series of lnet-selftests, the read and write rates range from 850MB/s - 875MB/s max.  I have performed the following optimizations to increase the data rate:
> 
> On the Client:
> lctl set_param osc.*.checksums=0
> lctl set_param osc.*.max_dirty_mb=256
> 
> On the OSTs
> lctl set_param obdfilter.*.writethrough_cache_enable=0
> lctl set_param obdfilter.*.read_cache_enable=0
> 
> echo 4096 > /sys/block/<devices>/queue/nr_requests
> 
> I have also loaded the ib_sdp module, which also brought an increase in speed.  However, we need to be able to record at no less than 1GB/s, which we cannot achieve right now.  Any thoughts on how I can optimize LNET, which clearly seems to be the bottleneck?
> 
> Thank you for any help you can provide,
> Carl Barberi
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org<mailto:Lustre-discuss at lists.lustre.org>
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
> 
> 
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org<mailto:Lustre-discuss at lists.lustre.org>
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
> 
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org<mailto:Lustre-discuss at lists.lustre.org>
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
> 
> 
> Confidentiality Notice: This e-mail message, its contents and any attachments to it are confidential to the intended recipient, and may contain information that is privileged and/or exempt from disclosure under applicable law. If you are not the intended recipient, please immediately notify the sender and destroy the original e-mail message and any attachments (and any copies that may have been made) from your system or otherwise. Any unauthorized use, copying, disclosure or distribution of this information is strictly prohibited. Email addresses that end with a "-c" identify the sender as a Fusion-io contractor.
>   
> 