Re: [Moosefs-users] Best FileSystem for chunkers.

[Steve W. <st...@pu...>]

On 05/17/2012 04:17 PM, Steve Wilson wrote:
> On 05/17/2012 04:05 PM, Atom Powers wrote:
>> On 05/17/2012 12:44 PM, Steve Wilson wrote:
>>> On 05/17/2012 03:26 PM, Atom Powers wrote:
>>>> * Compression, 1.16x in my environment
>>> I don't know if 1.16x would give me much improvement in performance.
>>> I typically see about 1.4x on my ZFS backup servers which made me
>>> think that this reduction in disk I/O could result in improved
>>> overall performance for MooseFS.
>> Not for performance, for disk efficiency. Ostensibly those 64MiB chunks
>> won't always use 64MiB with compression on, especially for smaller
>> files.
>
> This is a good point and it might help where it's most needed: all 
> those small configuration files, etc. that have a large impact on the 
> user's perception of disk performance.
>
>>>> Bad: * high RAM requirement
>>> Is the high RAM due to using raidz{2-3}?  I was thinking of making
>>> each disk a separate ZFS volume and then letting MooseFS combine the
>>> disks into an MFS volume (i.e., no raidz).  I realize that greater
>>> performance could be achieved by striping across disks in the chunk
>>> servers but I'm willing to trade off that performance gain for
>>> higher redundancy (in the case of using simple striping) and/or
>>> greater capacity (in the case of using raidz, raidz2, or raidz3).
>> ZFS does a lot of caching in RAM. My chunk servers use hardware RAID,
>> not raidz, and still use several hundred MiB of RAM.
>>
>> Personally, I would prefer to use raidz for muliple disks over MooseFS,
>> because managing individual disks and disk failures should be much
>> better. For example, to minimize the amount of re-balancing MooseFS
>> needs to do; not to mention the possible performance benefit. But I can
>> think of no reason why you couldn't do a combination of both.
>>
>
> That is certainly worth considering.  I hope to have enough time with 
> the new chunk servers to try out different configurations before I 
> have to put them into service.
>
> Steve