Unsolved
This post is more than 5 years old
25 Posts
0
2413
October 10th, 2013 12:00
Isilon Performance
I have a customer who has tested a simplistic single threaded test and wants to find out if the numbers make sense.
The dd test was with one dd process from one HPC node to the Isilon, so we expect when multiple instances are run that we will see the aggregate bandwidth improve. We just want to make sure that we are getting what is expected from this particular test.
The HPC and Isilon are connected via 10GB and the Onefs is shared via NFSv3.
A simple sequential write test with dd
For example, the following will write 1MB at a time to a file of 1GB size:
dd bs=1M count=1024 if=/dev/zero of= /file_1GB conv=fdatasync
I run this with different bs values (1K, 1M, 10M) and different file sizes (1GB, 10GB) to different file systems (local HD, isilon OneFS home and scrarch, NFS to Medusa frontend, Lustre in DOD, NFS to Matrix frontend, Matrix nas-0-0, Matrix nas-1-0) from a compute node. Here are the results:
|
bs=1K |
bs=1M |
bs=10M |
|||
File System |
file_1GB |
file_10GB |
file_1GB |
file_10GB |
file_1GB |
file_10GB |
Local HD |
|
100 MB/s |
|
|||
isilon OneFS (home) |
233 MB/s |
229 MB/s |
284 MB/s |
279 MB/s |
290 MB/s |
276 MB/s |
isilon OneFS (scratch) |
144 MB/s |
139 MB/s |
287 MB/s |
269 MB/s |
270 MB/s |
269 MB/s |
NFS to Medusa (10Gb) |
416 MB/s |
439 MB/s |
483 MB/s |
618 MB/s |
441 MB/s |
602 MB/s |
Lustre in DOD |
132 MB/s |
130 MB/s |
571 MB/s |
642 MB/s |
516 MB/s |
576 MB/s |
NFS to Matrix (1Gb) |
|
108 MB/s |
103 MB/s |
108 MB/s |
||
Matrix (nas-0-0) |
|
103 MB/s |
|
103 MB/s |
102 MB/s |
|
Matrix (nas-1-0) |
|
108 MB/s |
102 MB/s |
108 MB/s |
||
varunmalhi
25 Posts
0
October 11th, 2013 01:00
Hello Merk,
This is a closed network. There is no other traffic on this network. OneFS version is 7.0.1.5 and the nodes are X type.
Regards,
Varun Malhi
Peter_Sero
4 Operator
•
1.2K Posts
0
October 11th, 2013 01:00
As a reference, some actual figures on OneFS performance were given at the EMC World 2013:
https://www.emcworldonline.com/2013/connect/sessionDetail.ww?SESSION_ID=1589
The slides are downloadable as PDF, link on the right side.
(Have you considered the impression the dd table will make when adding,
as another baseline after "local HD", some throughput figures of, say, a
recent MacBook Air with local SSD?)
-- Peter
varunmalhi
25 Posts
0
October 11th, 2013 07:00
X400s to be specific
varunmalhi
25 Posts
0
October 14th, 2013 05:00
Merk,
This is for (redacted) and has research data which is essentially a lot of small files. At this point these are just test files. There are some large files but mostly they are small.
No there is no time requirements and remember, we are just talking about what kind of numbers can we expect? and why are the numbers off for home directories for small files as the results of the test show?
These are some questions that the customer needs to get answered. Is this normal performance expectation? Are there any published metrics for such tests for Isilon X Nodes?
Regards,
Varun Malhi