I'd be curious how things change with multiple io threads, as dd isn't necessarily a good performance investigation tool (you're rather testing "dd performance" as opposed to "using dd to test performance") if the concern is what to expect for your multi-tenant vm block store.
Personally, I get more bugged out over many-thread random read throughput or synchronous write latency.
On Friday, September 20, 2013, Jason Villalta wrote:
Thanks Jamie,I tried that too. But similar results. The issue looks to possibly be with the latency but everything is running on one server so logiclly I would think there would be no latency but according to this there may be something that is causing slow results. See Co-ResidencyI have not found a way to prove this to be true other than testing many difference configurations of OSDs and drives. At one point I had 3 OSDs all running one SSD drive. The performance was the same as when three OSDs were running on 3 separate SSD drives. Seems like there is something else going on here.Also I ran iotop while running rados bench and virtual machine sqlio. Write max out at 200-300MBps for the duration of the test. Reads never hit a sustained rate anywhere near that speed.
On Fri, Sep 20, 2013 at 7:18 PM, Jamie Alquiza <ja@xxxxxxxxxxxxxxxxx> wrote:--
I thought I'd just throw this in there, as I've been following this thread: dd also has an 'iflag' directive just like the 'oflag'.I don't have a deep, offhand recollection of the caching mechanisms at play here, but assuming you want a solid synchronous / non-cached read, you should probably specify 'iflag=direct'.
On Friday, September 20, 2013, Jason Villalta wrote:Mike,So I do have to ask, where would the extra latency be coming from if all my OSDs are on the same machine that my test VM is running on? I have tried every SSD tweak in the book. The primary concerning issue I see is with Read performance of sequential IOs in the 4-8K range. I would expect those to pull from three SSD disks on a local machine atleast as fast one Native SDD test. But I don't see that, its actually slower.On Wed, Sep 18, 2013 at 4:02 PM, Jason Villalta <jason@xxxxxxxxxxxx> wrote:
Thank Mike,High hopes right ;)I guess we are not doing too bad compared to you numbers then. Just wish the gap was a little closer between native and ceph per osd.
C:\Program Files (x86)\SQLIO>sqlio -kW -t8 -s30 -o8 -fsequential -b1024 -BH -LSc:\TestFile.datsqlio v1.5.SGusing system counter for latency timings, 100000000 counts per second8 threads writing for 30 secs to file c:\TestFile.datusing 1024KB sequential IOsenabling multiple I/Os per thread with 8 outstandingbuffering set to use hardware disk cache (but not file cache)using current size: 10240 MB for file: c:\TestFile.datinitialization doneCUMULATIVE DATA:throughput metrics:IOs/sec: 180.20MBs/sec: 180.20latency metrics:Min_Latency(ms): 39Avg_Latency(ms): 352Max_Latency(ms): 692histogram:ms: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24+%: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 100On Wed, Sep 18, 2013 at 3:55 PM, Mike Lowe <j.michael.lowe@xxxxxxxxx> wrote:
Well, in a word, yes. You really expect a network replicated storage system in user space to be comparable to direct attached ssd storage? For what it's worth, I've got a pile of regular spinning rust, this is what my cluster will do inside a vm with rbd writeback caching on. As you can see, latency is everything.dd if=/dev/zero of=1g bs=1M count=10241024+0 records in1024+0 records out1073741824 bytes (1.1 GB) copied, 6.26289 s, 171 MB/sdd if=/dev/zero of=1g bs=1M count=1024 oflag=dsync1024+0 records in1024+0 records out1073741824 bytes (1.1 GB) copied, 37.4144 s, 28.7 MB/sAs you can see, latency is a killer.On Sep 18, 2013, at 3:23 PM, Jason Villalta <jason@xxxxxxxxxxxx> wrote:Any other thoughts on this thread guys. I am just crazy to want near native SSD performance on a small SSD cluster?On Wed, Sep 18, 2013 at 8:21 AM, Jason Villalta <jason@xxxxxxxxxxxx> wrote:
That dd give me this.dd if=ddbenchfile of=- bs=8K | dd if=- of=/dev/null bs=8K8192000000 bytes (8.2 GB) copied, 31.1807 s, 263 MB/s
Which makes sense because the SSD is running as SATA 2 which should give 3Gbps or ~300MBpsI am still trying to better understand the speed difference between the small block speeds seen with dd vs the same small object size with rados. It is not a difference
--
-ja. Sent via mobile.
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
