Robert Olson | 1 Nov 2006 22:40
Favicon

PVFS and 2.6.18

Hi -

Looking at moving my PPC nodes to 2.6.18; it appears that the 1.5.1  
has problems (I googled & saw the note on the dev list about the  
issue in aug); I'm wondering what the right path is for upgrading.

Thanks,
--bob
Robert Latham | 1 Nov 2006 23:59
Favicon

Re: PVFS and 2.6.18

On Wed, Nov 01, 2006 at 03:40:35PM -0600, Robert Olson wrote:
> Looking at moving my PPC nodes to 2.6.18; it appears that the 1.5.1  
> has problems (I googled & saw the note on the dev list about the  
> issue in aug); I'm wondering what the right path is for upgrading.

Hi Bob

We could probably backport the patches for 2.6.18 to 1.5.1.  Looking
quickly, I can find at least three, but there could be others.   

We're close to a new release, though.  I think the plan is to release
around SC, so if you can hold off on the upgrade unti then, that might
be the easiest option.   

==rob

--

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Lab, IL USA                 B29D F333 664A 4280 315B
Robert Olson | 2 Nov 2006 00:04
Favicon

Re: PVFS and 2.6.18

Cool, that works for me.

Thanks,
--bob

On Nov 1, 2006, at 4:59 PM, Robert Latham wrote:

> On Wed, Nov 01, 2006 at 03:40:35PM -0600, Robert Olson wrote:
>> Looking at moving my PPC nodes to 2.6.18; it appears that the 1.5.1
>> has problems (I googled & saw the note on the dev list about the
>> issue in aug); I'm wondering what the right path is for upgrading.
>
> Hi Bob
>
> We could probably backport the patches for 2.6.18 to 1.5.1.  Looking
> quickly, I can find at least three, but there could be others.
>
> We're close to a new release, though.  I think the plan is to release
> around SC, so if you can hold off on the upgrade unti then, that might
> be the easiest option.
>
> ==rob
>
> -- 
> Rob Latham
> Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
> Argonne National Lab, IL USA                 B29D F333 664A 4280 315B
>
Andrew Pochinsky | 1 Nov 2006 23:06
Picon
Favicon

Growing storage

Hi,
	I'm looking to build a new PVFS2 storage and it seems that the price  
sweet spot these days is on boxes with a lot of local disks (my  
estimates are 5 to 10 TB per chassis). To provide enough bandwidth to  
clients, I'm thinking about connecting PVFS2 servers to the switch  
via 10Ge. (We already have a Cisco switch and can not go IB. GigE and  
10G are the only choices, as far as I know.) Did anybody have any  
experience with such a setup? What pitfalls are there?
Thanks,
--andrew
Dean Hildebrand | 2 Nov 2006 17:15
Picon
Favicon

Re: Growing storage

We are using Cisco and Force10 10Gb switches here at Michigan to do the 
opposite, 10Gb clients accessing many  1Gb storage nodes.  One thing 
we've found (and it is worse with our Cisco switch) is that the switch 
can run out of buffer space when packets switch from the 1Gb ports to 
the 10Gb port, and packets are dropped.  1Gb to 1Gb is good, 10Gb to 
10Gb is good, but 1Gb to 10Gb doesn't work so well (this is our read 
throughput benchmark).  The best throughput we've been able to get is 
around 4Gb/s in this scenario.  We are currently working with Force10 to 
try and resolve this issue, but no progress yet.

Hope this helps,

--

-- 
Dean Hildebrand
Ph.D. Candidate
University of Michigan

Andrew Pochinsky wrote:
> Hi,
>     I'm looking to build a new PVFS2 storage and it seems that the 
> price sweet spot these days is on boxes with a lot of local disks (my 
> estimates are 5 to 10 TB per chassis). To provide enough bandwidth to 
> clients, I'm thinking about connecting PVFS2 servers to the switch via 
> 10Ge. (We already have a Cisco switch and can not go IB. GigE and 10G 
> are the only choices, as far as I know.) Did anybody have any 
> experience with such a setup? What pitfalls are there?
> Thanks,
> --andrew
>
> _______________________________________________
(Continue reading)

Robert Olson | 2 Nov 2006 17:59
Favicon

PVFS2 & Jumbo Frames

The nodes in my cluster have dual Gig-E nics (tigon3). I'm pondering  
the installation of a second gigE switch to attach them all with a  
private storage network, and run it with jumbo frames. Do you think  
I'd see a significant performance increase with PVFS in this  
configuration?

We're currently running the nodes with standard 1500 byte frames as  
the nodes are sitting on a shared network with the rest of our group  
infrastructure.

I'm also curious if folks have experimented with the effects of the  
TCP segment offload stuff that is available on these NICs.

thanks,
--bob
Robert Latham | 2 Nov 2006 18:37
Favicon

Re: PVFS2 & Jumbo Frames

On Thu, Nov 02, 2006 at 10:59:51AM -0600, Robert Olson wrote:
> The nodes in my cluster have dual Gig-E nics (tigon3). I'm pondering  
> the installation of a second gigE switch to attach them all with a  
> private storage network, and run it with jumbo frames. Do you think  
> I'd see a significant performance increase with PVFS in this  
> configuration?

Hi Bob

We definitely expect jumbo frames to help out pvfs a lot.  Ti's had
good results on the teragrid, for example [1].  Unfortunately, I don't
think anyone has quantified the benefit.   

> I'm also curious if folks have experimented with the effects of the  
> TCP segment offload stuff that is available on these NICs.

Again, should help PVFS a lot, but we've got no hard data.  

==rob

[1] http://www.teragridforum.org/mediawiki/index.php?title=TG-11

--

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Lab, IL USA                 B29D F333 664A 4280 315B
Robert Olson | 2 Nov 2006 18:45
Favicon

Re: PVFS2 & Jumbo Frames

Thanks - interesting document!

--bob

On Nov 2, 2006, at 11:37 AM, Robert Latham wrote:

> On Thu, Nov 02, 2006 at 10:59:51AM -0600, Robert Olson wrote:
>> The nodes in my cluster have dual Gig-E nics (tigon3). I'm pondering
>> the installation of a second gigE switch to attach them all with a
>> private storage network, and run it with jumbo frames. Do you think
>> I'd see a significant performance increase with PVFS in this
>> configuration?
>
> Hi Bob
>
> We definitely expect jumbo frames to help out pvfs a lot.  Ti's had
> good results on the teragrid, for example [1].  Unfortunately, I don't
> think anyone has quantified the benefit.
>
>> I'm also curious if folks have experimented with the effects of the
>> TCP segment offload stuff that is available on these NICs.
>
> Again, should help PVFS a lot, but we've got no hard data.
>
> ==rob
>
> [1] http://www.teragridforum.org/mediawiki/index.php?title=TG-11
>
> -- 
> Rob Latham
(Continue reading)

Peng Gu | 2 Nov 2006 19:54
Picon

PVFS2 Cluster configuration

Dear all,

We are considering purchasing a 16-32 nodes cluster to serve as a
testbed for PVFS2 with GigaE or Infiniband interconnection. Do you
have any suggestions on
the configuration of the computer nodes and inter-connection hardware?

Thanks in advance!

Peng
Peng Gu | 3 Nov 2006 04:26
Picon

Re: PVFS2 Cluster configuration

In addition, we would like know your ideas about the following questions:

1) Whether 64-bit processor like Interl Xeon works fine with
PVFS2/future version or not.
2) Will Dell power-edge server cluster works fine?
3) Which OS, Linux 2.6 or other fit PVFS2/future version well.
4) Which vendor/cluster solution do you recommend?

Any inputs are appreciated,
Peng

On 11/2/06, Peng Gu <penggu <at> cs.ucf.edu> wrote:
> Dear all,
>
> We are considering purchasing a 16-32 nodes cluster to serve as a
> testbed for PVFS2 with GigaE or Infiniband interconnection. Do you
> have any suggestions on
> the configuration of the computer nodes and inter-connection hardware?
>
> Thanks in advance!
>
> Peng
>

Gmane