On 4/26/2016 9:57 AM, Chuck Lever wrote:
On Apr 26, 2016, at 10:13 AM, Steve Wise <swise@xxxxxxxxxxxxxxxxxxxxx> wrote:
Hey Chuck, I'm testing this series on cxgb4. I'm running 'iozone -a -+d -I' on a share and watching the server stats. Are the starve numbers expected?
Yes, unless you're seeing much higher numbers than
you used to.
Every 5.0s: for s in /proc/sys/sunrpc/svc_rdma/rdma_* ; do echo -n "$(basename $s): "; cat $s; done Tue Apr 26 07:10:17 2016
rdma_stat_read: 379872
rdma_stat_recv: 498144
rdma_stat_rq_poll: 0
rdma_stat_rq_prod: 0
rdma_stat_rq_starve: 675564
This means work was enqueued on the svc_xprt, but by the
time the upper layer invoked svc_rdma_recvfrom, the work
was already handled by an earlier wake-up.
I'm not exactly sure why this happens, but it seems to be
normal (if suboptimal).
rdma_stat_sq_poll: 0
rdma_stat_sq_prod: 0
rdma_stat_sq_starve: 1748000
No SQ space to post a Send, so the caller is put to sleep.
The server chronically underestimates the SQ depth, especially
for FRWR. I haven't figured out a better way to estimate it.
But it's generally harmless, as there is a mechanism to put
callers to sleep until there is space on the SQ.
Thanks.
With this iw_cxgb4 drain fix applied:
[PATCH 3/3] iw_cxgb4: handing draining an idle qp
http://www.spinics.net/lists/linux-rdma/msg34927.html
The series tests good over cxgb4.
Tested-by: Steve Wise <swise@xxxxxxxxxxxxxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
