[CentOS] 40TB File System Recommendations

[Ross Walker]

On Apr 15, 2011, at 12:32 PM, Rudi Ahlers <Rudi at SoftDux.com> wrote:

> 
> 
> On Fri, Apr 15, 2011 at 6:26 PM, Ross Walker <rswwalker at gmail.com> wrote:
> On Apr 15, 2011, at 9:17 AM, Rudi Ahlers <Rudi at SoftDux.com> wrote:
> 
>> 
>> 
>> On Fri, Apr 15, 2011 at 3:05 PM, Christopher Chan <christopher.chan at bradbury.edu.hk> wrote:
>> On Friday, April 15, 2011 07:24 PM, Benjamin Franz wrote:
>> > On 04/14/2011 09:00 PM, Christopher Chan wrote:
>> >>
>> >> Wanna try that again with 64MB of cache only and tell us whether there
>> >> is a difference in performance?
>> >>
>> >> There is a reason why 3ware 85xx cards were complete rubbish when used
>> >> for raid5 and which led to the 95xx/96xx series.
>> >> _
>> >
>> > I don't happen to have any systems I can test with the 1.5TB drives
>> > without controller cache right now, but I have a system with some old
>> > 500GB drives  (which are about half as fast as the 1.5TB drives in
>> > individual sustained I/O throughput) attached directly to onboard SATA
>> > ports in a 8 x RAID6 with *no* controller cache at all. The machine has
>> > 16GB of RAM and bonnie++ therefore used 32GB of data for the test.
>> >
>> > Version  1.96       ------Sequential Output------ --Sequential Input-
>> > --Random-
>> > Concurrency   1     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
>> > --Seeks--
>> > Machine        Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP
>> > /sec %CP
>> > pbox3        32160M   389  98 76709  22 91071  26  2209  95 264892  26
>> > 590.5  11
>> > Latency             24190us    1244ms    1580ms   60411us   69901us
>> > 42586us
>> > Version  1.96       ------Sequential Create------ --------Random
>> > Create--------
>> > pbox3               -Create-- --Read--- -Delete-- -Create-- --Read---
>> > -Delete--
>> >                 files  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP
>> > /sec %CP
>> >                    16 10910  31 +++++ +++ +++++ +++ 29293  80 +++++ +++
>> > +++++ +++
>> > Latency               775us     610us     979us     740us     370us
>> > 380us
>> >
>> > Given that the underlaying drives are effectively something like half as
>> > fast as the drives in the other test, the results are quite comparable.
>> 
>> Woohoo, next we will be seeing md raid6 also giving comparable results
>> if that is the case. I am not the only person on this list that thinks
>> cache is king for raid5/6 on hardware raid boards and the using hardware
>> raid + bbu cache for better performance one of the two reasons why we
>> don't do md raid5/6.
>> 
>> 
>> >
>> > Cache doesn't make a lot of difference when you quickly write a lot more
>> > data than the cache can hold. The limiting factor becomes the slowest
>> > component - usually the drives themselves. Cache isn't magic performance
>> > pixie dust. It helps in certain use cases and is nearly irrelevant in
>> > others.
>> >
>> 
>> Yeah, you are right - but cache is primarily to buffer the writes for
>> performance. Why else go through the expense of getting bbu cache? So
>> what happens when you tweak bonnie a bit?
>> _______________________________________________
>> 
>> 
>> 
>> As matter of interest, does anyone know how to use an SSD drive for cach purposes on Linux software RAID  drives? ZFS has this feature and it makes a helluva difference to a storage server's performance. 
> 
> Put the file system's log device on it.
> 
> -Ross
> 
> 
> _______________________________________________
> 
> 
> 
> Well, ZFS has a separate ZIL for that purpose, and the ZIL adds extra protection / redundancy to the whole pool. 
> 
> But the Cache / L2ARC drive caches all common reads & writes (simply put) onto SSD to improve overall system performance. 
> 
> So I was wondering if one could do this with mdraid or even just EXT3 / EXT4?

Ext3/4 and XFS allow specifying an external log device which if is an SSD can speed up writes. All these file systems aggressively use page cache for read/write cache. The only thing you don't get is L2ARC type cache, but I heard of a dm-cache project that might provide provide that type of cache.

-Ross

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.centos.org/pipermail/centos/attachments/20110415/413018fa/attachment.html>