Re: HDD vs SSD without explanation - Mailing list pgsql-performance

On 16/01/18 23:14, Neto pr wrote:
> 2018-01-15 20:04 GMT-08:00 Mark Kirkwood <mark.kirkwood@catalyst.net.nz>:
>> On 16/01/18 13:18, Fernando Hevia wrote:
>>
>>>
>>>
>>> The 6 Gb/s interface is capable of a maximum throughput of around 600
>>> Mb/s. None of your drives can achieve that so I don't think you are limited
>>> to the interface speed. The 12 Gb/s interface speed advantage kicks in when
>>> there are several drives installed and it won't make a diference in a single
>>> drive or even a two drive system.
>>>
>>> But don't take my word for it. Test your drives throughput with the
>>> command Justin suggested so you know exactly what each drive is capable of:
>>>
>>>      Can you reproduce the speed difference using dd ?
>>>      time sudo dd if=/dev/sdX of=/dev/null bs=1M count=32K
>>>      skip=$((128*$RANDOM/32)) # set bs to optimal_io_size
>>>
>>>
>>> While common sense says SSD drive should outperform the mechanical one,
>>> your test scenario (large volume sequential reads) evens out the field a
>>> lot. Still I would have expected somewhat similar results in the outcome, so
>>> yes, it is weird that the SAS drive doubles the SSD performance. That is why
>>> I think there must be something else going on during your tests on the SSD
>>> server. It can also be that the SSD isn't working properly or you are
>>> running an suboptimal OS+server+controller configuration for the drive.
>>>
>> I would second the analysis above - unless you see your read MB/s slammed up
>> against 580-600MB/s contunuously then the interface speed is not the issue.
>> We have some similar servers that we replaced 12x SAS with 1x SATA 6 GBit/s
>> (Intel DC S3710) SSD...and the latter way outperforms the original 12 SAS
>> drives.
>>
>> I suspect the problem is the particular SSD you have - I have benchmarked
>> the 256GB EVO variant and was underwhelmed by the performance. These
>> (budget) triple cell nand SSD seem to have highly variable read and write
>> performance (the write is all about when the SLC nand cache gets
>> full)...read I'm not so sure of - but it could be crappy chipset/firmware
>> combination. In short I'd recommend *not* using that particular SSD for a
>> database workload. I'd recommend one of the Intel Datacenter DC range (FWIW
>> I'm not affiliated with Intel in any way...but their DC stuff works well).
>>
>> regards
>>
>> Mark
> Hi Mark
> In other forums one person said me that on samsung evo should be
> partition aligned to 3072 not  default 2048 , to start on erase block
> bounduary .  And fs block should be 8kb. I am studing this too. Some
> DBAs have reported in other situations that the SSDs when they are
> full, are very slow. Mine is 85% full, so maybe that is also
> influencing. I'm disappointed with this SSD from Samsung, because in
> theory, the read speed of an SSD should be more than 300 times faster
> than an HDD and this is not happening.
>
>

Interesting - I didn't try changing the alignment. However I could get 
the rated write and read performance on simple benchmarks (provided it 
was in a PCIe V3 slot)...so figured it was ok with the default aligning. 
However once more complex workloads were attempted (databases and 
distributed object store) the performance was disappointing.

If the SSD is 85% full that will not help either (also look at the 
expected lifetime of these EVO's - not that great for a server)!

One thing worth trying is messing about with the IO scheduler: if you 
are using noop, then try deadline (like I said crappy firmware)...

Realistically, I'd recommend getting an enterprise/DC SSD (put the EVO 
in your workstation, it will be quite nice there)!

Cheers
Mark