The iflag addition should help with at least having more accurate reads via
dd, but in terms of actually testing performance, have you tried sysbench
or bonie++?

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-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 <[email protected]>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 <[email protected]>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 <[email protected]>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 <[email protected]> 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 <[email protected]>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
>
> --
> --
> *Jason Villalta*
> Co-founder
> [image: Inline image 1]
> 800.799.4407x1230 | www.RubixTechnology.com<http://www.rubixtechnology.com/>
>


-- 
-ja. Sent via mobile.

<<EmailLogo.png>>

_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com

Reply via email to