[rsyslog] [OT] Re: reliability of SSD disks?

david at lang.hm david at lang.hm
Thu Aug 20 05:35:28 CEST 2009


On Wed, 19 Aug 2009, Mr. Demeanour wrote:

> Aaron Wiebe wrote:
>> SSD's are as reliable, if not more reliable, than your regular
>> spinning rust.
>>
>> But if you want to get even more speed, and just as much reliability,
>>  check out fusionio.com.  They're launching a consumer market PCIe
>> card at 80GB this year (at $800 a card).  The technology is -very-
>> cool though, I got a presentation about it this week (and met the
>> woz!).
>
> So the website (http://www.fusionio.com) doesn't seem to offer any
> information on the number of write cycles these devices are able to
> sustain. Since they are NAND Flash, this would be expected to be up to
> about a million, rendering them of dubious value as a part of a logging
> system. With wear-levelling, once cells start dying, I'd expect the
> death-rate to climb very rapidly indeed.
>
> I would have thought that DRAM-based devices would be more suitable in
> this role.

remember that with wear leveling, that is a million writes to each spot on 
the drive. even with the 'write magnification' effect (where every write 
that actually goes to disk requires doing an entire eraseblock, on a 80G 
drive with 128K blocks that is 640,000,000,000 writes, or if doing the 
theortical max of 100,000/sec it will only last 6,400,000 seconds, or 74 
days.

if only writing at 10,000 logs/sec that becomes 740 days or two years

if you go with a 160G drive of the same specs the wearout numbers double.

another option is to go with a raid card with battery backed cache, if 
that cache can delay the writes to the flash drive itself so that instead 
of a fsync for every message causing a write for every 256 characters it 
does it for every 256K the wearout time for the flash drives skyrockets 
(74,000 days, or 200 years)

Intel claims that their drives only have a write magnification factor of 
~1.2:1 or so. to do this they would have to do some caching on the drive. 
if they really do achieve this sort of result, the lifetime of their 
drives would be measured in years or decades of service at max write rates

yes, for extremely high traffic volumes it may be better to go with 
something like the ANS-9010 Serial ATA RAM disk 
http://techreport.com/articles.x/16255/1 but it's significantly more 
expensive, and it requires a 5.25 drive bay (hard to get in rackmount 
equipment nowdays), but it is even faster than the flash SSDs

David Lang



More information about the rsyslog mailing list