Re: higth performance write to disk

On Wed, Sep 4, 2013 at 2:19 PM, Tomas Vondra <tv(at)fuzzy(dot)cz> wrote:
> On 4.9.2013 20:52, Jeison Bedoya Delgado wrote:
>> Hi merlin, Thanks for your interest, I'm using version 9.2.2, I have
>> a machine with 128GB RAM, 32 cores, and my BD weighs 400GB. When I
>> say slow I meant that while consultations take or failing a copy with
>> pgdump, this taking the same time before had 10K disks in raid 10 and
>> now that I have SSDs in Raid 10.
>>
>> That behavior is normal, or you can improve writing and reading.
>
> SSDs are great random I/O, not that great for sequential I/O (better
> than spinning drives, but you'll often run into other bottlenecks, for
> example CPU).
>
> I'd bet this is what you're seeing. pg_dump is a heavily sequential
> workload (read the whole table from start to end, write a huge dump to
> the disk). A good RAID array with 10k SAS drives can give you very good
> performance (I'd say ~500MB/s reads and writes for 6 drives in RAID10).
> I don't think the pg_dump will produce the data much faster.
>
> Have you done any tests (e.g. using fio) to test the performance of the
> two configurations? There might be some hw issue but if you have no
> benchmarks it's difficult to judge.
>
> Can you run the fio tests now? The code is here:
>
> http://freecode.com/projects/fio
>
> and there are even a basic example:
> http://git.kernel.dk/?p=fio.git;a=blob_plain;f=examples/ssd-test.fio
>
>
> And how exactly are you running the pg_dump? And collect some basic
> stats next time it's running, for example a few samples from
>
> vmstat 5
> iostat -x -k 5
>
> and watch top how much CPU it's using.

yeah. Also, some basic stats would be nice. For example how much data
is getting written out and how long is it taking? We need to
establish benchmark of 'slow'.

merlin