[mythtv-users] MySQL on SSD 99% Utilization

Raymond Wagner raymond at wagnerrp.com
Thu Feb 3 18:51:09 UTC 2011


On 2/3/2011 13:46, Raymond Wagner wrote:
> On 2/3/2011 13:38, Matthew McClement wrote:
>> On 02/03/2011 05:43 PM, Raymond Wagner wrote:
>>> On 2/3/2011 12:37, Matthew McClement wrote:
>>>> On 02/03/2011 05:27 PM, Raymond Wagner wrote:
>>>>> On 2/3/2011 11:40, Brian Long wrote:
>>>>>> The CPUs are doing almost nothing so it appears the MySQL updates
>>>>>> being performed during a commflag are screwing with the SSD.  Is it
>>>>>> generally a bad practice to put MySQL on a SSD?  I figured it would
>>>>>> provide a nice fast DB and keep my DB separate from my recording
>>>>>> disks, but I guess I was wrong.
>>>>> Looking at reviews on the SSDNow V 30GB, the one thing that stands 
>>>>> out
>>>>> is awful random read/write performance.  Even still, it should be far
>>>>> better than a disk drive is capable of.
>>>> Anands benchmarks show the SSDNow V 30GB is the same speed as a
>>>> Velociraptor at 4K random writes, which is pretty terrible for a SSD:
>>>>
>>>> http://www.anandtech.com/bench/Product/155?vs=182
>>> That metric just looks wrong.  It looks like they're buffering 
>>> writes to
>>> something other than the drive, and that something else is actually 
>>> what
>>> is bottlenecking.  That's the only way to account for the discrepancy
>>> between reads and writes.  Measuring write buffer speed is meaningless
>>> unless that write buffer is battery backed and recoverable.
>> Eh? SSD's are significantly slower at writes than reads, especially if
>> you're writing to a non-GC'd cell, which with how the Toshiba controller
>> handles TRIM is entirely possible(GC isn't continuous but rather seems
>> to be demand triggered). If you then add non-aligned sub-page sized
>> writes, things can get bad pretty quickly.
>>
>> It's buffering and other tricks which make SSD's faster today over the
>> early versions, rather than slower.
>>
>> And it's not like that result is an oddity. Just look at any benchmark
>> for the early JMicron based SSDs to see SSDs that were often *slower*
>> than hard disks at random writes.
> That benchmark showed write performance on the velociraptor to be ~20x 
> that of the read performance.

Make that ~3x.



More information about the mythtv-users mailing list