cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Heath Oderman <he...@526valley.com>
Subject Re: New User: OSX vs. Debian on Cassandra 0.5.0 with Thrift
Date Fri, 16 Apr 2010 15:36:29 GMT
Ok, it took me a long time to get py_stress working.

I didn't have thrift / boost / gcc on my debian box :)  ....

I'm using this command line believing it's similar to my c# tests from a
remote box:

cnb:~/apache-cassandra-0.6.0-src/contrib/py_stress# python stress.py -o
insert -n 1000000 -d 10.113.0.195 -y super

its wicked fast, as you'd expect.

total,interval_op_rate,avg_latency,elapsed_time
61765,6176,0.00804715166591,10
126667,6490,0.00763260502319,20
184952,5828,0.00798357854098,30
249697,6474,0.00813590502387,40
310190,6049,0.00820316256729,50
370394,6020,0.00823702464729,60
431557,6116,0.00811347469409,70
492085,6052,0.00818783322119,80
558895,6681,0.00740800772112,90
620426,6153,0.00805101232533,100
685168,6474,0.00766180823714,110
748768,6360,0.00779282277485,120
811008,6224,0.00797537056523,130
867327,5631,0.00881408287019,140

This leaves me further stumped.  I guess i will try running py_stress from a
remote box, because I've got to believe it has something to do with the
connection.

Thanks for nudging me toward py_stress.  I'm no closer to understanding, but
I have more info!

Stu

On Thu, Apr 15, 2010 at 1:57 PM, Jonathan Ellis <jbellis@gmail.com> wrote:

> What kind of numbers do you get from contrib/py_stress?
>
> (that's located somewhere else in 0.5, but you should really be using
> 0.6 anyway.)
>
> On Thu, Apr 15, 2010 at 12:53 PM, Heath Oderman <heath@526valley.com>
> wrote:
> > So checking it out quickly:
> > vmstat -
> > Never swaps.  si and so  stay at 0 during the load.
> > iostat -x
> > the %util never climbs above 0.00, but the avgrg-sz jumps bewteen samples
> > from 0 - 30 - 90 - 0 (5 second intervals)
> > top shows the cpu barely working and mem utilization is below 20%.
> > Still slow.  :(
> > Thanks for the suggestions.  In your article on your blog it'd be awesome
> to
> > include some implications, like "avgrg-sz over 250 may mean XXX"  Even if
> > it's utterly hardware and system dependent it'd give a guy like me an
> idea
> > if what I was seeing was bad or good. :D
> > Thanks again,
> > Heath
> >
> > On Thu, Apr 15, 2010 at 1:34 PM, Heath Oderman <heath@526valley.com>
> wrote:
> >>
> >> Thanks Jonathan, I'll check this out right away.
> >>
> >> On Thu, Apr 15, 2010 at 1:32 PM, Jonathan Ellis <jbellis@gmail.com>
> wrote:
> >>>
> >>> You're right, to get those numbers on debian something is very wrong.
> >>>
> >>> Have you looked at
> >>> http://spyced.blogspot.com/2010/01/linux-performance-basics.html ?
> >>> What is the bottleneck on the linux machines?
> >>>
> >>> With the kind of speed you are seeing I wouldn't be surprised if it is
> >>> swapping.
> >>>
> >>> -Jonathan
> >>>
> >>> On Tue, Apr 13, 2010 at 11:38 PM, Heath Oderman <heath@526valley.com>
> >>> wrote:
> >>> > Hi,
> >>> > I wrote a few days ago and got a few good suggestions.  I'm still
> >>> > seeing
> >>> > dramatic differences between Cassandra 0.5.0 on OSX vs. Debian Linux.
> >>> > I've tried on Debian with the Sun JRE and the Open JDK with nearly
> >>> > identical
> >>> > results. I've tried a mix of hardware.
> >>> > Attached are some graphs I've produced of my results which show that
> in
> >>> > OSX,
> >>> > Cassandra takes longer with a greater load but is wicked fast
> >>> > (expected).
> >>> > In the SunJDK or Open JDK on Debian I get amazingly consistent time
> >>> > taken to
> >>> > do the writes, regardless of the load and the times are always
> >>> > ridiculously
> >>> > high.  It's insanely slow.
> >>> > I genuinely believe that I must be doing something very wrong in my
> >>> > Debian
> >>> > setups, but they are all vanilla installs, both 64 bit and 32 bit
> >>> > machines,
> >>> > 64bit and 32 bit installs.  Cassandra packs taken from
> >>> > http://www.apache.org/dist/cassandra/debian.
> >>> > I am using Thrift, and I'm using a c# client because that's how I
> >>> > intend to
> >>> > actually use Cassandra and it seems pretty sensible.
> >>> > An example of what I'm seeing is:
> >>> > 5 Threads Each writing 100,000 Simple Entries
> >>> > OSX: 1 min 16 seconds ~ 6515 Entries / second
> >>> > Debian: 1 hour 15 seconds ~ 138 Records / second
> >>> > 15 Threads Each writing 100,000 Simple Entries
> >>> > OSX: 2min 30 seconds seconds writing ~10,000 Entries / second
> >>> > Debian: 1 hour 1.5 minutes ~406 Entries / second
> >>> > 20 Threads Each Writing 100,000 Simple Entries
> >>> > OSX: 3min 19 seconds ~ 10,050 Entries / second
> >>> > Debian: 1 hour 20 seconds ~ 492 Entries / second
> >>> > If anyone has any suggestions or pointers I'd be glad to hear them.
> >>> > Thanks,
> >>> > Stu
> >>> > Attached:
> >>> > 1. CassLoadTesting.ods (all my results and graphs in OpenOffice
> format
> >>> > downloaded from Google Docs)
> >>> > 2. OSX Records per Second - a graph of how many entries get written
> per
> >>> > second for 10,000 & 100,000 entries as thread count is increased
in
> >>> > OSX.
> >>> > 3. Open JDK Records per Second - the same graph but of Open JDK on
> >>> > Debian
> >>> > 4. Open JDK Total Time By Thread - the total time taken from test
> start
> >>> > to
> >>> > finish (all threads completed) to write 10,000 & 100,000 entries
as
> >>> > thread
> >>> > count is increased in Debian with Open JDK
> >>> > 5. OSX Total time by Thread - same as 4, but for OSX.
> >>> >
> >>> >
> >>
> >
> >
>

Mime
View raw message