[Ksummit-2008-discuss] Suggested topic: possible

Eyal Shani Eyal.Shani at sandisk.com
Thu Aug 7 07:49:31 PDT 2008


Chris Mason & Matthew Wilcox wrote:

>>
>> Oh ... something I should point out.  If you're doing 4k IOs, you need a
>> drive capable of 75000 IOPS.  If you're doing 128k IOs, it only has to
>> be capable of 2400 IOPS.
>>

>I was mostly trying to underline the internal parallelism of the SSD.
>I'd expect 32 contiguous 4k reads o be slower than a single 128k because
>the SSD can really read it in parallel internally, and the big IO gives
>it the chance to optimize things better.

>This matters with reads even if the ssd becomes very good at doing tiny
>writes.

What you state is correct - internal parallelism of SSD is among the important design issues a good SSD put effort into perfecting.
As the SSDs are becoming better & better in internal Read/Write latency, the protocol overhead of multiple small transactions vs. single long transaction becomes increasingly more painful, and will become the bottleneck.

Eyal Shani @ SanDisk SSD


More information about the Ksummit-2008-discuss mailing list