Re: WAL in RAM

From: Merlin Moncure <mmoncure(at)gmail(dot)com>
To: Tomas Vondra <tv(at)fuzzy(dot)cz>
Cc: Marcus Engene <mengpg2(at)engene(dot)se>, pgsql-performance(at)postgresql(dot)org
Subject: Re: WAL in RAM
Date: 2011-10-28 18:40:17
Message-ID: CAHyXU0whWGCmEJjM1xFFMpjPkhVbKnLcYrnAKPFK3jt1BFKScQ@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-performance

On Fri, Oct 28, 2011 at 1:26 PM, Tomas Vondra <tv(at)fuzzy(dot)cz> wrote:
> On 28 Říjen 2011, 18:11, Merlin Moncure wrote:
>> On Fri, Oct 28, 2011 at 10:28 AM, Marcus Engene <mengpg2(at)engene(dot)se> wrote:
>>> Hi list,
>>>
>>> Every now and then I have write peaks which causes annoying delay on my
>>> website. No particular reason it seems, just that laws of probability
>>> dictates that there will be peaks every now and then.
>>>
>>> Anyway, thinking of ways to make the peaks more bareable, I saw the new
>>> 9.1
>>> feature to bypass WAL. Problems is mainly that some statistics tables
>>> ("x
>>> users clicked this link this month") clog the write cache, not more
>>> important writes. I could live with restoring a nightly dump of these
>>> tables
>>> and loose a days worth of logs.
>>>
>>> Though not keen on jumping over to early major versions an old idea of
>>> putting WAL in RAM came back. Not RAM in main memory but some thingie
>>> pretending to be a drive with proper battery backup.
>>>
>>> a) It seems to exist odd hardware with RAM modules and if lucky also
>>> battery
>>> b) Some drive manufactureres have done hybird ram-spindle drives
>>> (compare
>>> with possibly more common ssd-spindle hybrides).
>>>
>>> b) sounds slightly more appealing since it basically means I put
>>> everything
>>> on those drives and it magically is faster. The a) alternatives also
>>> seemed
>>> to be non ECC which is a no-no and disturbing.
>>>
>>> Does anyone here have any recommendations here?
>>>
>>> Pricing is not very important but reliability is.
>>
>> Have you ruled out SSD?  They are a little new, but I'd be looking at
>> the Intel 710.  In every case I've seen SSD permanently ends I/O
>> issues.  DRAM storage solutions I find to be pricey and complicated
>> when there are so many workable flash options out now.
>
> Are you sure SSDs are a reasonable option for WAL? I personally don't
> think it's a good option, because WAL is written in a sequential manner,
> and that's not an area where SSDs beat spinners really badly.
>
> For example the Intel 710 SSD has a sequential write speed of 210MB/s,
> while a simple SATA 7.2k drive can write about 50-100 MB/s for less than
> 1/10 of the 710 price.
>
> I'm not saying SSDs are a bad thing, but I think it's a waste of money to
> use them for WAL.

sure, but then you have to have a more complicated setup with a
drive(s) designated for WAL, another for storage, etc. Also, your
argument falls away if the WAL is shared with another drive. The era
of the SSD is here. All new systems I plan will have SSD storage
unless cost pressures are extreme -- often with a single drive unless
you need the extra storage. If I need availability, instead of RAID,
I'll just build hot standby in.

merlin

In response to

Responses

Browse pgsql-performance by date

  From Date Subject
Next Message Robert Haas 2011-10-28 19:03:14 Re: should i expected performance degradation over time
Previous Message Tomas Vondra 2011-10-28 18:26:19 Re: WAL in RAM