[lustre-discuss] Small sequential reads on cached file are slow

[Hans Henrik Happe]

Thanks a lot. I'm looking forward to test this. Wounder why I haven't stumpled across this issue. While HPC workloads try to avoid syscall overhead, there are lots of software that rely on the OS to cope with small IOs. Think the average user will become a lot happier.

Cheers,
Hans Henrik

On October 2, 2016 11:21:55 PM GMT+02:00, "Dilger, Andreas" <andreas.dilger at intel.com> wrote:
>Please test with Lustre master (pre-2.9.0). There were optimizations
>landed specifically to improve small file read performance.  While
>Lustre 2.9.0 isn't released yet, it is getting very close. 
>
>Cheers, Andreas
>
>> On Oct 2, 2016, at 06:54, Hans Henrik Happe <happe at nbi.ku.dk> wrote:
>> 
>> Hi,
>> 
>> While testing small sequential reads I noticed that Lustre is more
>than 10 times slower than local fs when reading cached data. So
>basically no network I/O to OSSes. I wanted to check if readahead was
>working for this small I/O case, but it seems that cached case isn't.
>> 
>> The client is running Lustre 2.8 and CentOS 6.8.
>> 
>> Lustre:
>> 
>> $ dd if=file of=/dev/null bs=512
>> 2097152+0 records in
>> 2097152+0 records out
>> 1073741824 bytes (1.1 GB) copied, 20.5081 s, 52.4 MB/s
>> 
>> $ dd if=file of=/dev/null bs=4k
>> 262144+0 records in
>> 262144+0 records out
>> 1073741824 bytes (1.1 GB) copied, 2.91177 s, 369 MB/s
>> 
>> $ dd if=file of=/dev/null bs=1M
>> 1024+0 records in
>> 1024+0 records out
>> 1073741824 bytes (1.1 GB) copied, 0.160732 s, 6.7 GB/s
>> 
>> Local fs:
>> 
>> $ dd if=/tmp/file of=/dev/null bs=512
>> 2097152+0 records in
>> 2097152+0 records out
>> 1073741824 bytes (1.1 GB) copied, 1.56432 s, 686 MB/s
>> 
>> $ dd if=/tmp/file of=/dev/null bs=4k
>> 262144+0 records in
>> 262144+0 records out
>> 1073741824 bytes (1.1 GB) copied, 0.275451 s, 3.9 GB/s
>> 
>> $ dd if=/tmp/file of=/dev/null bs=1M
>> 1024+0 records in
>> 1024+0 records out
>> 1073741824 bytes (1.1 GB) copied, 0.148798 s, 7.2 GB/s
>> 
>> 
>> Is this a known issue?
>> 
>> Cheers,
>> Hans Henrik Happe
>> _______________________________________________
>> lustre-discuss mailing list
>> lustre-discuss at lists.lustre.org
>> http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org

-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20161004/50d30fd4/attachment.htm>