[rhelv6-list] Distributed parallel fault-tolerant file systems

Prentice Bisbal prentice at ias.edu
Mon Mar 5 22:47:12 UTC 2012


On 03/05/2012 05:30 PM, Phil Meyer wrote:
> On 03/05/2012 02:28 PM, Prentice Bisbal wrote:
>> On 03/05/2012 04:11 PM, Phil Meyer wrote:
>>> On 03/05/2012 09:34 AM, Bohmer, Andre ten wrote:
>>>> Hello,
>>>>
>>>> Any advise or experience from production systems regarding distributed
>>>> parallel fault-tolerant file systems like Lustre ?
>>>> We would like to offer high performance, redundant storage via nfs
>>>> and cifs
>>>> from Redhat servers.
>>>>
>>>> At this we've HP's XP9000 Ibrix in use, but performance is not all
>>>> that
>>>> great.
>>>>
>>> My test results are nearly 2 years old by now, but at that time we
>>> concluded that, for NFS, the very best performing and low
>>> administration solution was Isilon.
>>>
>>> Need more space?  Drop in a storage module and its online in~60
>>> seconds.  Performance starting to lag a bit under load? Drop in a CPU
>>> module and its online in ~60 seconds.
>>>
>>> More expensive than any home grown, for sure, but well worth it in my
>>> opinion.
>>>
>>> Two years now in a very heavily used environment without an issue.
>>
>> NFS is not a parallel filesystem, and it's not fault-tolerant, which
>> were the OP's requirements. NFS is fine for most day-to-day use (that's
>> why it's so ubiquitous), but when you have 100+ nodes accessing the
>> server at once, you can see it's weaknesses, which is why filesystems
>> like Lustre and PVFS started appeariing soon after Linux clusters came
>> onto the scene.
>>
>
> NFS on Isilon servers IS parallel, and IS fault tolerant.  The
> connection is a single connection, yes, but underneath it is
> distributed, and they offer several methods to help load balance the
> clients.
>
> I don't mean to sell Isilon here, but it is very good for what it does.

I'm not doubting that - I've heard good things about Isilon.

Unless your Isilon is using pNFS, the *NFS* is not parallel. The data
access within the Isilon between the head unit and the disks might be
parallel, but clients accessing the data on the Isilon are not doing it
in parallel, therefore the NFS itself is not parallel. And this is the
bottleneck that parallel filesystems are designed to deal with.

Prentice




More information about the rhelv6-list mailing list