Subject: Re: Distributing data over "spindles" even on AWS EBS, (followup to
 the work queue saga)
To: Jeremy Schneider <schnjere@amazon.com>,
 pgsql-performance@lists.postgresql.org
References: <b49c6a04-caaf-a3d1-4f07-978558260cf8@gusw.net>
 <679375f7-bb61-6a66-e91a-4b5454f9223b@gusw.net>
 <b2648191-36b1-9ef2-5833-24b1c8d1cd26@amazon.com>
From: Gunther <raj@gusw.net>
Message-ID: <74d58c5d-2e56-781d-4bbc-b122ce3e6020@gusw.net>
Date: Sun, 17 Mar 2019 14:42:04 -0400
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:60.0) Gecko/20100101
 Thunderbird/60.5.1
MIME-Version: 1.0
In-Reply-To: <b2648191-36b1-9ef2-5833-24b1c8d1cd26@amazon.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Content-Language: en-US
Precedence: bulk

On 3/14/2019 11:11, Jeremy Schneider wrote:
> On 3/14/19 07:53, Gunther wrote:
>>   2. build a low level "spreading" scheme which is to take the partial
>>      files 4653828 and 4653828.1, .2, _fsm, etc. and move each to another
>>      device and then symlink it back to that directory (I come back to this!)
> ...
>> To 2. I find that it would be a nice feature of PostgreSQL if we could
>> just use symlinks and a symlink rule, for example, when PostgreSQL finds
>> that 4653828 is in fact a symlink to /otherdisk/PG/16284/4653828, then
>> it would
>>
>>   * by default also create  4653828.1 as a symlink and place the actual
>>     data file on /otherdisk/PG/16284/4653828.1
> How about if we could just specify multiple tablespaces for an object,
> and then PostgreSQL would round-robin new segments across the presently
> configured tablespaces?  This seems like a simple and elegant solution
> to me.

Very good idea! I agree.

Very important also would be to take out the existing patch someone had 
contributed to allow toast tables to be assigned to different tablespaces.

>>   4. maybe I can configure in AWS EBS to reserve more IOPS -- but why
>>      would I pay for more IOPS if my cost is by volume size? I can just
>>      make another volume? or does AWS play a similar trick on us with
>>      IOPS being limited on some "credit" system???
> Not credits, but if you're using gp2 volumes then pay close attention to
> how burst balance works. A single large volume is the same price as two
> striped volumes at half size -- but the striped volumes will have double
> the burst speed and take twice as long to refill the burst balance.

Yes, I learned that too. It seems a very interesting "bug" of the Amazon 
GP2 IOPS allocation scheme.  They say it's like 3 IOPS per GiB, so if I 
have 100 GiB I get 300 IOPS. But it also says minimum 100. So that means 
if I have 10 volumes of 10 GiB each, I get 1000 IOPS minimum between 
them all. But if I have it all on one 100 GiB volume I only get 300 IOPS.

I wonder if Amazon is aware of this. I hope they are and think that's 
just fine. Because I like it.

It also is a clear sign to me that I want to use page sizes > 4k for the 
file system. I have tried on Amazon Linux to use 8k block sizes of the 
XFS volume, but I cannot mount those, since the Linux says it can 
currently only deal with 4k blocks. This is another reason I consider 
switching the database server(s) to FreeBSD.  OTOH, who knows may be 
this 4k is a limit of the AWS EBS infrastructure. After all, if I am 
scraping the 300 or 1000 IOPS limit already and if I can suddenly 
upgrade my block sizes per IO, I double my IO throughput.

regards,
-Gunther