While running a test to improve read performance on 1GbE and 10GbE
clients, a memory leak in the async readahead code was discovered. A
trivial fix follows.
In case it is of interest, using three separately hosted OSDs (10GbE),
each configured to use six disks, i get the following read performance
(single threaded read, using find $DIR -type f -exec dd bs=8M if={}
of=/dev/null, file system has 1.5TiB of data in variously sized objects)
v3.0.0 (1GbE client): 500Mb/s
v3.0.0 (10GbE client): 1Gb/s
v3.0.0 + ceph-client/master (1GbE client): >900Mb/s
v3.0.0 + ceph-client/master (10GbE client): ~2.5-3Gb/s
Of note:
- doubling the read ahead window for the 10GbE client didn't buy
any more performance (default = 8192*1024).
- adding more threads to do reads on the 10GbE client only brought
aggregate throughput up to <4Gbit/sec
- using 8 1GbE clients, an aggregate throughput of ~6Gbit/s is achieved
- increasing to 12GbE clients, aggregate increases only up to 6.5Gbit/sec
( I shall retry this with a different disk configuration for the OSDs )
The numbers above are fairly woolly, but hopefully illustrate current
throughput improvements from the ceph-client/master branch.
Kind regards,
..david
David Flynn (1):
ceph: fix memory leak in async readpages
fs/ceph/addr.c | 1 +
1 files changed, 1 insertions(+), 0 deletions(-)
--
1.7.4.1
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
