[Beowulf] NFSoIB and SRP benchmarking

Vincent Diepeveen diep at xs4all.nl
Mon Aug 27 04:38:20 PDT 2012


On Aug 27, 2012, at 1:31 PM, Vincent Diepeveen wrote:

> Yeah it will be a great machines. You get what you pay for of course.
> Confirming that is not even interesting. If you pay big bucks  
> obviously it'll be fast.
>
> For such machines, just break the raid card during operation, then  
> replace it with another
> card and see whether your data still is there and whether it can  
> work without too much problems
> within the crucial 2 minutes that you've got normally at exchanges  
> to get new equipment in
> (not counting rebuild time), or you're fired.
>
> So for rich financials i bet such machines are attractive.
>
> As for HPC, scaling is important.
>
> As for cheap scaling....
>
> If i search for benchmarks of raid controllers, you can find  
> everything as well.
>
> What really misses is benchmarks of built in raid controllers.
>
> If you want to scale cheap it's interesting to know how to do the i/ 
> o obviously.
>
> The motherboards i've got have the ESB2 from intel built in as RAID  
> controller.
>
> If i google i see nothing on expected read/write speeds.
>
> Nearly all raid controllers you always see have a limit of 1 GB/s,  
> that really sucks
> compared to even the bandwidth one can generate to the RAM.
>
> Basically 1 GB/s with 8 cores Xeon 2.5Ghz means you've got 8 *  
> 2.5Ghz = 20k cycles a byte, or 32k instructions you can
> execute for reading or writing just 1 byte.

Oh dear calculations on monday... ...make that 20Ghz * 4 instructions  
per cycle / 1 GB = 80 cycles a byte...

Or for each 'bitboard' which is 64 bits, so similar to a double, if  
you read and write it.

You read 2 and write 1.

That's 24 bytes of i/o.

So that's 80 cycles * 24 = 1920 instructions. I didn't factor in  
SSE4.2 yet...

For i/o intensive applications bandwidth is interesting to have to  
the i/o. I don't see how file servers deliver that for clusters.


>
> Most raid controllers that are cheap, say a 100-300 euro (130 - 400  
> dollar), they're not managing more than a
> write speed of a 600MB/s , that's in fact in raid-0, not really  
> realistic, and 700-900MB/s read speed.
>
> That's with around a drive or 8.
>
> We know however that when the raid array is empty, those drives  
> will handsdown get a write speed at outside of drives of
> a 130MB/s, so the limitation is the raid cards CPU itself. Usually  
> around a 800Mhz at the better cards.
>
> Many benchmark with SAS drives. I see some fujitsu drives that are  
> rated 188MB/s at outer side of drives, which benchmark to
> 900MB/s readspeed.
>
> The limitation seems to be the raidcards in most cases.
>
> So my simple question would be: is this better with the built in  
> raid controllers that most motherboards have as they can use
> the fast 2.5Ghz Xeon cpu's?
>
> On Aug 26, 2012, at 12:22 AM, holway at th.physik.uni-frankfurt.de wrote:
>
>> Hello Beowulf 2.0
>>
>> I've just started playing with NFSoIB to provide super fast backend
>> storage for a bunch of databases that we look after here. Oracle and
>> Nexenta are both sending me ZFS based boxes to test and I hope to  
>> compare
>> the performance and stability of these with the Netapp (formally lsi
>> engenio) E5400.
>>
>> This will be the first time I will be getting into serious storage
>> benchmarking. Does anyone have any interesting tests they would  
>> like to
>> run or and experience performing these kinds of tests?
>>
>> I have 4 HP G8 boxes with 96GB ram each QDR IB and 10G ethernet as  
>> consumers.
>>
>> I will also be testing the performance of KVM and Virtuozzo  
>> (commercial
>> version of OpenVZ) which is a kernel sharing virtualization  
>> similar to BSD
>> Jails.
>>
>> Thanks,
>>
>> Andrew
>>
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
>> Computing
>> To change your subscription (digest mode or unsubscribe) visit  
>> http://www.beowulf.org/mailman/listinfo/beowulf
>



More information about the Beowulf mailing list