Re: SSDs with Postgresql? - Mailing list pgsql-general

On 29/04/11 16:35, Greg Smith wrote:
> On 04/26/2011 10:30 AM, Toby Corkindale wrote:
>> I see Intel is/was claiming their SLC SSDs had a *minimum* lifetime of
>> 2PB in writes for their 64GB disks; for your customer with a 50GB db
>> and 20GB/day of WAL, that would work out at a minimum lifetime of a
>> million days, or about 273 years!
>> The cheaper "consumer grade" MLC drives should still last minimum 5
>> years at 20GB/day according to their literature. (And what I found was
>> fairly out of date)
>> That doesn't seem too bad to me - I don't think I've worked anywhere
>> that keeps their traditional spinning disks in service beyond 5 years
>> either.
>
>
> The comment I made there was that the 20GB/day system was a very small
> customer. One busy server, who are also the ones most likely to want
> SSD, I just watched recently chug through 16MB of WAL every 3
> seconds=450GB/day. Now, you're right that those systems also aren't
> running with a tiny amount of flash, either. But the write volume scales
> along with the size, too. If you're heavily updating records in
> particular, the WAL volume can be huge relative to the drive space
> needed to store the result.

For that sort of workload, it does sound like SSD isn't ideal.
Although if the Intel figure of 2PB is to be believed, it'd still take
 >200 years to wear the drives out (assuming you'll need at least ten 64
GB disks just to store the DB on).
(Reference: http://download.intel.com/design/flash/nand/extreme/319984.pdf )

But yeah, much sooner on the cheaper MLC drives, as I understand it..
seems harder to get info out of Intel on their max write amounts.

> As for the idea that I'm just singling out one anecdote, I have
> terabytes of lost data on multiple systems behind my negativity here. I
> was just pointing out a public failure that included some post-mortem I
> liked. I'm not sure if I have any happy customers who were early
> adopters of regular SLC or MLC drives really; the disaster rate is very
> close to 100% for the first few generations of those drives I've seen,
> and I've been around 50-ish of them. I'm hoping the current models
> shipping now are better, getting the write cache stuff sorted out better
> will be a big help. But it's been a scary technology for database use so
> far. The published numbers from the manufacturer literature are a very
> rosy best case when you're hitting the disk with this type of workload.

Ah, thanks - it's interesting to hear more of your experiences there.

I do note that more recent SSD drives have made many improvements for
durability - now some of them are doing a sort of internal-RAID over
their NAND chips, so that if/when bits of them die, you don't lose your
data, and can keep operating.
(Reference: http://www.anandtech.com/show/4244/intel-ssd-320-review/2 )

It sounds like the technology is more mature now, but I guess we won't
know until more people have been using it, successfully, for a while..

-Toby