Hi, I have setup Lustre with 12 OSTs, 1 MDS and 3 Clients. I intend to test Lustre performance against some extreme workloads. To get my initial results I wanted to try accessing a Large number of small files (Which will lead to a lot of MDS load) and then try access of a small number of Large files (Which will be more of Client-OST interactions) With these prilimenary tests I want to use the results to give an Idea of Lustre performance against extreme cases. But after reseraching tools for this I am not sure how to really simulate these workloads and the input charecterstics to be able to compare the outputs and give a believable result. Can you suggest a way to accomplish the above? Thank You Priya -------------- next part -------------- An HTML attachment was scrubbed... URL: http://mail.clusterfs.com/pipermail/lustre-discuss/attachments/20060719/7cb14bd9/attachment.html
Priya,
What tools have you looked at? I wanted a tool which would manage
different
workloads simultaneously and have the capability to barrier on a
per-block basis.
I ended up writing something like this a few months ago. It''s still a
bit crufty
since I haven''t had time to clean it up. It does support mpi or
threads..
Here''s an example config file which runs 4pes, 2 which create and 2
which write.
If you think this would be useful for you let me know and I''ll put the
source
on the web.
paul
group 2peCreateUnlink {
files_per_dir = 200000;
tree_depth = 0;
tree_width = 0;
pes = 2;
test_freq = 0;
path = /lustre-300sc/pauln/;
output_path = /lustre-300sc/pauln/;
filename = 2peCreateUnlink;
file_size = 0;
block_size = 0;
thrash_lock = no;
samedir = no;
fsync_block = no;
verify = no;
block_barrier = no;
app_barrier = no;
time_barrier = no;
iterations = 1;
iotests (
createOnly [create]
unlink [unlink]
)
}
group 2peWriter {
files_per_dir = 2;
tree_depth = 0;
tree_width = 0;
pes = 2;
test_freq = 0;
path = /lustre-300sc/pauln/;
output_path = /lustre-300sc/pauln/;
filename = 2peWriter_1MBb;
file_size = 8g;
block_size = 1m;
thrash_lock = no;
samedir = yes;
fsync_block = no;
time_block = yes;
verify = no;
block_barrier = no;
app_barrier = no;
time_barrier = no;
iterations = 1;
iotests (
write00 [create:openwr:write:close]
unlinkOnly [unlink]
)
}
Priya Muralidharan wrote:
> Hi,
>
> I have setup Lustre with 12 OSTs, 1 MDS and 3 Clients. I intend to
> test Lustre performance against some extreme workloads.
> To get my initial results I wanted to try accessing a Large number of
> small files (Which will lead to a lot of MDS load) and then try access
> of a small number of Large files (Which will be more of Client-OST
> interactions)
>
> With these prilimenary tests I want to use the results to give an
> Idea of Lustre performance against extreme cases. But after
> reseraching tools for this I am not sure how to really simulate these
> workloads and the
> input charecterstics to be able to compare the outputs and give a
> believable result.
>
> Can you suggest a way to accomplish the above?
>
> Thank You
> Priya
>
>
>------------------------------------------------------------------------
>
>_______________________________________________
>Lustre-discuss mailing list
>Lustre-discuss@clusterfs.com
>https://mail.clusterfs.com/mailman/listinfo/lustre-discuss
>
My MDS load-test was:
echo
{0,1,2,3,4,5,6,7,8,9}{0,1,2,3,4,5,6,7,8,9}{0,1,2,3,4,5,6,7,8,9}{0,1,2,3,4,5,6,7,8,9}{0,1,2,3,4,5,6,7,8,9}
| xargs touch
and then deleting them. This creates 100 000 0-byte files in a single
directory.
As for client tests, a compilation job like linux kernel or
OpenOffice.org do it quite nicely.
However my recommendation is trying something based on what you intend
to use lustre for.
Best regards,
Jo?o Miguel Neves
Qua, 2006-07-19 ?s 16:31 -0700, Priya Muralidharan
escreveu:> Hi,
>
> I have setup Lustre with 12 OSTs, 1 MDS and 3 Clients. I intend to
> test Lustre performance against some extreme workloads.
> To get my initial results I wanted to try accessing a Large number of
> small files (Which will lead to a lot of MDS load) and then try access
> of a small number of Large files (Which will be more of Client-OST
> interactions)
>
> With these prilimenary tests I want to use the results to give an
> Idea of Lustre performance against extreme cases. But after
> reseraching tools for this I am not sure how to really simulate these
> workloads and the
> input charecterstics to be able to compare the outputs and give a
> believable result.
>
> Can you suggest a way to accomplish the above?
>
> Thank You
> Priya
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss@clusterfs.com
> https://mail.clusterfs.com/mailman/listinfo/lustre-discuss
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Esta =?ISO-8859-1?Q?=E9?= uma parte de mensagem
assinada digitalmente
Url :
http://mail.clusterfs.com/pipermail/lustre-discuss/attachments/20060723/2f493f5e/attachment.bin
----- Original Message ----- From: "Jo?o Miguel Neves" <joao@silvaneves.org> To: "Priya Muralidharan" <Priya_Muralidharan@symantec.com> Cc: <lustre-discuss@clusterfs.com> Sent: Sunday, July 23, 2006 8:45 AM Subject: Re: [Lustre-discuss] Lustre Performance Analysis> _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss@clusterfs.com > https://mail.clusterfs.com/mailman/listinfo/lustre-discuss >Priya, Due to the way Lustre works, I would recommend that the small files contain at least one byte of data. The work required to create zero sized files is different than the work required to create a file that contains any data. Zero sized files is a special case in Lustre, and it is doubtful that you actually have a real-world workload the creates 100,000 zero sized files. As Neves points out, the closer your test is to what you actually intend to do, is very important in the evaluation.... Enjoy, Don Capps