Phil Schwan wrote:>On Tue, 2004-09-21 at 07:48, Martin Vogt wrote: > > >>I have tested lustre 1.4beta with kernel 2.4.24 over NFS with IOZone. >> >>iozone -i 0 -i 1 -b ./iozone-2.xls -L 32 -M -r 16k -s 2048000 -S 512 -f >>/a/file >> >>Here are some results: >> >>1. Local NFS System performance (no lustre) MB/s >>2. iozone normal lustre performance (no NFS) MB/s >>3. iozone in lustre-nfs share MB/s (locally mounted) >> >>Test 3 is the "lustre over nfs" test which, it is a bit slow... >> >>1: Write1: 40 Write2: 52 Read1: 30 Read2: 35 >>2: Write1: 108 Write2: 107 Read1: 65 Read2: 65 >>3: Write1: 4 Write2: 4 Read1: 1 Read2: 1 <- !!! >> >> >>Here a test with two clients and a remote NFS mount: >>(aggregated throughput) >> >>4: Write1: 4 Write2: 3 Read1: 4 Read: 4 <- !!! >> >> > >Exactly which version of Lustre are you running? These numbers don''t >reflect our testing, and it makes me suspect that you are accidentally >running the wrong version, because of some confusing numbering. > >-Phil > >Hello, I have used: lustre-b1-4_mballoc-20040714.tar.gz The throughput is really slow, so it can be an error in my testing setup, but it was a simple /etc/export entry and a mount command, the same as in "test 1" using the local disk. regards, Martin
Hi Martin-- On 9/23/2004 5:47, "Martin Vogt" <vogt@itwm.fraunhofer.de> wrote:> I have used: > > lustre-b1-4_mballoc-20040714.tar.gz > > The throughput is really slow, so it can be an error in my testing setup, > but it was a simple /etc/export entry and a mount command, the same > as in "test 1" using the local disk.I checked into this, and your problem is now well-understood. First, that 2004-07-14 version of Lustre is older than the 1.3.x beta snapshots on the web site; I will remove this older snapshot, because it''s more likely to cause confusion than provide any real value at this point. Secondly, I learned that the NFS optimizations which will be part of Lustre 1.4 have not been included in the 1.3.x snapshots to date -- but they will be included in the 1.3.3 beta snapshot, which should be available in a week or so. We would appreciate it if you tried again with the new snapshot and reported your results. Just to calibrate your expectations, Lustre re-exporting NFS will not be as fast as a native client, but it should be much better than the 1 to 4 MB/s that you are seeing currently. -Phil
On Tue, 2004-09-21 at 07:48, Martin Vogt wrote:> > I have tested lustre 1.4beta with kernel 2.4.24 over NFS with IOZone. > > iozone -i 0 -i 1 -b ./iozone-2.xls -L 32 -M -r 16k -s 2048000 -S 512 -f > /a/file > > Here are some results: > > 1. Local NFS System performance (no lustre) MB/s > 2. iozone normal lustre performance (no NFS) MB/s > 3. iozone in lustre-nfs share MB/s (locally mounted) > > Test 3 is the "lustre over nfs" test which, it is a bit slow... > > 1: Write1: 40 Write2: 52 Read1: 30 Read2: 35 > 2: Write1: 108 Write2: 107 Read1: 65 Read2: 65 > 3: Write1: 4 Write2: 4 Read1: 1 Read2: 1 <- !!! > > > Here a test with two clients and a remote NFS mount: > (aggregated throughput) > > 4: Write1: 4 Write2: 3 Read1: 4 Read: 4 <- !!!Exactly which version of Lustre are you running? These numbers don''t reflect our testing, and it makes me suspect that you are accidentally running the wrong version, because of some confusing numbering. -Phil
Hello,
some month ago, there was a thread about lustre over nfs:
>Phil Schwan * phil@clusterfs.com <mailto:phil%40clusterfs.com>/Thu,
13 May 2004 14:09:48 -0400:
>>On Fri, 2004-05-07 at 12:04, Dan Stromberg wrote:
>>Is anyone out there nfs exporting a lustre mount?
>>Are you using the patches from clusterfs that are supposed to speed up
>>this combination?
>>
>These patches aren''t available yet. Lustre 1.4.x will contain
many
>performance improvements, some of which are specifically for making an
>NFS export of a Lustre client perform reasonably.
>It''s possible that some people are doing NFS exports already, but
the
>code available today did not receive a lot of attention in this area.
>If they are, it is probably extremely slow, and not entirely stable.
Hello,
I have tested lustre 1.4beta with kernel 2.4.24 over NFS with IOZone.
iozone -i 0 -i 1 -b ./iozone-2.xls -L 32 -M -r 16k -s 2048000 -S 512 -f
/a/file
Here are some results:
1. Local NFS System performance (no lustre) MB/s
2. iozone normal lustre performance (no NFS) MB/s
3. iozone in lustre-nfs share MB/s (locally mounted)
Test 3 is the "lustre over nfs" test which, it is a bit slow...
1: Write1: 40 Write2: 52 Read1: 30 Read2: 35
2: Write1: 108 Write2: 107 Read1: 65 Read2: 65
3: Write1: 4 Write2: 4 Read1: 1 Read2: 1 <- !!!
Here a test with two clients and a remote NFS mount:
(aggregated throughput)
4: Write1: 4 Write2: 3 Read1: 4 Read: 4 <- !!!
regards,
Martin Vogt