Lustre discuss - Jul 2008 - Lustre Performance Data for Simultaneous Reads and Writes from Multiple Clients

I?m working with someone who is modeling a customer system, and wants to
partially model Lustre performance as part of that.

What they would like is the following data, or similar, for a given network.
I say given in that pick any network config and any given stripe size and
any given file size (can be 250MB to 1GB), and then supply the following
data:

* From a single client, the read I/O start and stop time, or ?bandwidth?
* From a single client, the write I/O start and stop time, or ?bandwidth?
* Then introduce additional clients doing reads or writes and study the
impact, as in one client writing and four clients reading simultaneously,
and their start/stop times for IO, or bandwidth, and then one client reading
and four clients writing simultaneously, and their individual IO bandwidths

The objective really is to know how concurrent reads and writes impact
Lustre performance.

Does this data exist, or would someone need to go do and collect this data?

Thanks,
Dan


-- 
Daniel Ferber
Lustre DOE, DOD, and NSF Partnerships
Sun Microsystems Federal
+1 612-486-5167 (Office)
+1 651-356-9481 (Cell)
Daniel.Ferber at Sun.com
Sun Internal: x29704


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.lustre.org/pipermail/lustre-discuss/attachments/20080725/29c339ef/attachment.html

Hi Daniel,

I don?t believe so.

Various people have posted informal results from their own tests in the
field, but none have ever been formally collated. There are some rough
numbers on the Wikipedia page for CFS for GigE, IB, and 10GigE, but they
assume particular things about configuration, disk throughput, striping, OST
counts, etc.

Because Lustre is so versatile, information like this can be hard to nail
down ? running a GigE network with ATA drives is obviously not going to get
the best performance compared to 8 GB Fibre Channel, but both are equally
valid Lustre configurations.

hth,
Klaus

On 7/25/08 9:03 AM, "Daniel Ferber" <Daniel.Ferber at
Sun.COM>did etch on stone
tablets:
> 
> I?m working with someone who is modeling a customer system, and wants to
> partially model Lustre performance as part of that.
> 
> What they would like is the following data, or similar, for a given
network. I
> say given in that pick any network config and any given stripe size and any
> given file size (can be 250MB to 1GB), and then supply the following data:
> 
> * From a single client, the read I/O start and stop time, or ?bandwidth?
> * From a single client, the write I/O start and stop time, or ?bandwidth?
> * Then introduce additional clients doing reads or writes and study the
> impact, as in one client writing and four clients reading simultaneously,
and
> their start/stop times for IO, or bandwidth, and then one client reading
and
> four clients writing simultaneously, and their individual IO bandwidths
> 
> The objective really is to know how concurrent reads and writes impact
Lustre
> performance. 
> 
> Does this data exist, or would someone need to go do and collect this data?
> 
> Thanks,
> Dan
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.lustre.org/pipermail/lustre-discuss/attachments/20080725/2f5c7df7/attachment-0001.html

Dan,

I have posted some results for our large Lustre FS deployments, focusing 
almost exclusively on writing large, global arrays from MPI jobs. I have 
both cases of realistic, large-scale usage (many processors, with 
smaller buffer sizes on each, respectively); to unrealistic scenarios 
(very large buffers for each processor, with a smaller processor count) 
to see the optimal condition, maximal byte-rate for our configuration 
and hardware.

Your interest area seems very different from ours. Can you give some 
more information about it? What are the counts for the additional 
clients and how are they managed? Are the multi-client executions run 
with MPI? Do you have a benchmark application you are working with, such 
as IOR?


Marty Barnaby




Klaus Steden wrote:
>
> Hi Daniel,
>
> I don?t believe so.
>
> Various people have posted informal results from their own tests in 
> the field, but none have ever been formally collated. There are some 
> rough numbers on the Wikipedia page for CFS for GigE, IB, and 10GigE, 
> but they assume particular things about configuration, disk 
> throughput, striping, OST counts, etc.
>
> Because Lustre is so versatile, information like this can be hard to 
> nail down ? running a GigE network with ATA drives is obviously not 
> going to get the best performance compared to 8 GB Fibre Channel, but 
> both are equally valid Lustre configurations.
>
> hth,
> Klaus
>
> On 7/25/08 9:03 AM, "Daniel Ferber" <Daniel.Ferber at
Sun.COM>did etch on
> stone tablets:
>
>
>     I?m working with someone who is modeling a customer system, and
>     wants to partially model Lustre performance as part of that.
>
>     What they would like is the following data, or similar, for a
>     given network. I say given in that pick any network config and any
>     given stripe size and any given file size (can be 250MB to 1GB),
>     and then supply the following data:
>
>         * From a single client, the read I/O start and stop time, or
>           ?bandwidth?
>         * From a single client, the write I/O start and stop time, or
>           ?bandwidth?
>         * Then introduce additional clients doing reads or writes and
>           study the impact, as in one client writing and four clients
>           reading simultaneously, and their start/stop times for IO,
>           or bandwidth, and then one client reading and four clients
>           writing simultaneously, and their individual IO bandwidths
>
>
>     The objective really is to know how concurrent reads and writes
>     impact Lustre performance.
>
>     Does this data exist, or would someone need to go do and collect
>     this data?
>
>     Thanks,
>     Dan
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.lustre.org/pipermail/lustre-discuss/attachments/20080728/72a724cb/attachment.html