On Sat, 12 Sep 2009, Doug Hughes wrote:

> [email protected] wrote:
>> On Sat, 12 Sep 2009, Doug Hughes wrote:
>> 
>>> [email protected] wrote:
>>>>> 
>>>>> 
>>>>> adding standard DRAM onto this with a super capacity for flushing the
>>>>> ram to nonvolatile flash gives you the best of all worlds. Its still
>>>>> relatively new and relatively expensive. I know of only one vendor using
>>>>> this in mainstream, and that's Sun with their Readzilla and Logzilla
>>>>> accelerators for FishWorks. (Other vendors may be doing similar, but I'm
>>>>> not aware of them)
>>>> 
>>>> although unless the drive includes battery backup for the DRAM you will 
>>>> loose a lot of data if the drive looses power.
>>> oops. Typo on my part. super 'capacitor', which acts as a battery in that 
>>> it flushes the DRAM to nonvolatile flash storage
>> 
>> writing to flash takes a considerable amount of power, do they really put 
>> capacitors that large on the boards? (I sure haven't seen anything that 
>> looked like a capacitor that large in any of the disassembled pictures of 
>> flash drives that I see in the reviews)
>> 
> yes: http://blogs.sun.com/ahl/entry/fishworks_launch
> But, remember, this is NOT a normal flash drive.

I had misunderstood you. I thought you were talking about flash drives, 
not storage arrays (storage arrays do frequently use DRAM, they have a 
full processor on board and large battery packs)

>>> zfs normally stripes the intent log across the data disks unless you tell 
>>> it otherwise (e.g. 46 disks on a Sun x45000 or x4540)
>>> You can also put the ZIL on a dedicated high speed rotating drive. The 
>>> x25E outperforms both except in the case where you want high sequential 
>>> throughput.
>> 
>> well given that the X25E is faster than the rotating storage I would expect 
>> it to be a win there as well.
>> 
>> but the point I was trying to make was that I suspect most of the benifit 
>> is in the step from a striped intent log to an intent log on dedicated 
>> media, with the move from there to an X25 helping, but not as much.
>> 
> I am trying to correct this misconception. I assure you it's much bigger than 
> that for 7200 RPM SATA to X-25E. The latency is lower and throughput is 
> higher. But, don't believe me. Check out the references I posted earlier.
> Even the consumer X25-M sata SSD is .1msec vs 4.2msec on a 7200RPM SATA 
> enterprise drive. for the X25-E it is .085msec for for 15KRPM SCSI it is 
> 3.9msec. (The fusion IO is .05 msec)
>
> The raw sequential throughput of the X-25E is close to 3000MB/sec on a 64KB 
> record whereas your 9200 RPM drive will be around 100 and the 15KRPM drive 
> will be aroudn 227. (it doubles on smaller block sizes)

I don't doubt your performance number (not much anyway, the X25E is 
limited to the speed of sata, so a theoretical limit of 300MB/sec, not 
3000MB/sec ;-), but the huge win (just like it is with databases) is when 
you move your intent log off of drives that are seeking all over the place 
to a drive where you don't have to seek much, if at all, letting your 
writes _be_ sequential in the first place

> I strongly encourage everybody interested in this topic to look at the CMU 
> paper on this.
> Included here again: 
> http://www.pdsi-scidac.org/events/PDSW08/resources/papers/simsa_PDSW.pdf
> Or just review the PPT slides if you don't want to read the whole paper:
> http://www.pdsi-scidac.org/events/PDSW08/resources/slides/simsa_PDSW.pdf
>
>
>>> I have evaluated a battery backed ram that was much slower than flash for 
>>> ZIL purposes. I will spare the vendor some embarassment, but I was very 
>>> surprised with the results and still don't fully understand why.
>> 
>> no question that you can find lemons in any technology, many raid cards 
>> have traditionally just tried to be faster than the drives that they ran 
>> (enough to show an improvement), but the good ones compete with each other 
>> on performance, so with any of them I would not expect that sort of 
>> problem.
>> 
> The weird thing was it was showing 400MB/sec on read/write throughput over SC 
> quite easily using iozone, but for ZIL offloading it was slower than having 
> the ZIL striped across the 7200 RPM sata drives with the data by a factor of 
> 2-4x
>
> So, it wasn't just a case of the DRAM device being generically slow. (the 
> vendor didn't necessarily understand it either)

interesting

David Lang
_______________________________________________
Discuss mailing list
[email protected]
http://lopsa.org/cgi-bin/mailman/listinfo/discuss
This list provided by the League of Professional System Administrators
 http://lopsa.org/

Reply via email to