Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Sat, May 9, 2009 at 6:14 AM, Lars Ellenberg <lars.ellenberg at linbit.com>wrote: > On Fri, May 08, 2009 at 06:27:36PM -0400, Gennadiy Nerubayev wrote: > > Possibly related to my earlier post, but this time I focused on two > things > > primarily: SSDs and metadata. > > > > Hardware: DRBD on top of 1 X25-e in each of the DRBD nodes, configured as > > simple volume on a hardware raid controller, and connected via IP over > > infiniband. > > Workload: 100% random io writes, 8KB block size > > > > Direct to the disk: 77.78MB/s, average latency 2ms > > Disconnected DRBD, metadata on ramdisk: 75.64MB/s, average latency 2ms > > Connected DRBD, metadata on ramdisk: 50.87MB/s, average latency 4ms > > Disconnected DRBD, metadata internal: 6.25MB/s, average latency 39ms > > Connected DRBD, metadata internal: 6.20MB/s, average latency 39ms > > Disconnected DRBD, metadata on a single 15K sas disk: 43.46MB/s, average > > latency 5ms > > Connected DRBD, metadata on a single 15K sas disk: 39.32MB/s, average > > latency 5ms > > Could you add a "dm linear" to the table? > i.e. just checking if just one small layer of "virtual" block device > has any effect on throughput and or latency. > dmsetup create experiment <<<"0 $[20 <<21] linear /dev/sdX" > then do your benchmark against /dev/mapper/experiment > (just to see if that performs any different than "raw" /dev/sdx) Doesn't seem to work: dmsetup create experiment <<< "0 $[20 <<21] linear /dev/sdb" device-mapper: reload ioctl failed: Invalid argument Command failed > > Full resync speeds for all tests were 180-200MB/s (about what's expected > > from a single x25-e). There is no difference between flexible and regular > > metadata for internal or external usage (metadata was recreated for those > > tests). Interestingly, ~6MBs is the same speed that I got when testing a > 7 > > disk raid 0 15K sas array with internal metadata (everything else is the > > same) and putting metadata on ramdisk moved that up to ~35MB/s. > > > > So for some reason even in the case of a very fast SSD, internal metadata > > performance for random writes is really bad. Putting it on any kind of > > external disk brings an immediate exponential performance increase.. > > IO scheduler? try deadline and noop. > echo deadline > /sys/block/sdX/queue/scheduler Oooh, aaah: Disconnected DRBD, metadata internal, deadline: 42.59MB/s, average latency 5ms Connected DRBD, metadata internal, deadline: 39.34MB/s, average latency 5ms Disconnected DRBD, metadata internal, noop: 44.06MB/s, average latency 5ms Connected DRBD, metadata internal, noop: 38.61MB/s, average latency 5ms After running the above deadline and noop tests, I went back to cfq and got identical results as in the original test (~6MB/s). Kernel for reference is 2.6.29.1. -Gennadiy -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.linbit.com/pipermail/drbd-user/attachments/20090509/8c6fd874/attachment.htm>