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-Residency http://ceph.com/docs/master/rados/troubleshooting/troubleshooting-osd/
I 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 <j...@grey-boundary.com> 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 <ja...@rubixnet.com>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 -LS >> c:\TestFile.dat >> sqlio v1.5.SG >> using system counter for latency timings, 100000000 counts per second >> 8 threads writing for 30 secs to file c:\TestFile.dat >> using 1024KB sequential IOs >> enabling multiple I/Os per thread with 8 outstanding >> buffering set to use hardware disk cache (but not file cache) >> using current size: 10240 MB for file: c:\TestFile.dat >> initialization done >> CUMULATIVE DATA: >> throughput metrics: >> IOs/sec: 180.20 >> MBs/sec: 180.20 >> latency metrics: >> Min_Latency(ms): 39 >> Avg_Latency(ms): 352 >> Max_Latency(ms): 692 >> histogram: >> 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 >> 0 100 >> >> >> >> On Wed, Sep 18, 2013 at 3:55 PM, Mike Lowe <j.michael.l...@gmail.com>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=1024 >> 1024+0 records in >> 1024+0 records out >> 1073741824 bytes (1.1 GB) copied, 6.26289 s, 171 MB/s >> dd if=/dev/zero of=1g bs=1M count=1024 oflag=dsync >> 1024+0 records in >> 1024+0 records out >> 1073741824 bytes (1.1 GB) copied, 37.4144 s, 28.7 MB/s >> >> As you can see, latency is a killer. >> >> On Sep 18, 2013, at 3:23 PM, Jason Villalta <ja...@rubixnet.com> 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 <ja...@rubixnet.com>wrote: >> >> That dd give me this. >> >> dd if=ddbenchfile of=- bs=8K | dd if=- of=/dev/null bs=8K >> 8192000000 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 ~300MBps >> >> I 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 of a few MB per sec. It seems to nearly be a >> factor of 10. I just want to know if this is a hard limit in Ceph or a >> factor of the underlying disk speed. Meaning if I use spindles to read >> data would the speed be the same or would the read speed be a factor of 10 >> less than the speed of the underlying disk? >> >> >> On Wed, Sep 18, 2013 at 4:27 AM, Alex Bligh <a...@alex.org.uk> wrote: >> >> >> On 17 Sep 2013, at 21:47, Jason Villalta wrote: >> >> > dd if=ddbenchfile of=/dev/null bs=8K >> > >> >> > > -- > -ja. Sent via mobile. > -- -- *Jason Villalta* Co-founder [image: Inline image 1] 800.799.4407x1230 | www.RubixTechnology.com<http://www.rubixtechnology.com/>
<<EmailLogo.png>>
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com