Subject: Re: [PATCH] xlogreader: do not read a file block twice
To: Arthur Zakirov <a.zakirov@postgrespro.ru>,
 pgsql-hackers@lists.postgresql.org
Cc: Grigory Smolkin <g.smolkin@postgrespro.ru>
References: <2ddf4a32-517e-d6f4-d992-4a63b6035bfd@postgrespro.ru>
From: Andrey Lepikhov <a.lepikhov@postgrespro.ru>
Organization: Postgres Professional
Message-ID: <93d74870-3217-7ab3-b220-31cab64b7c7a@postgrespro.ru>
Date: Tue, 12 Feb 2019 22:47:32 +0500
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101
 Thunderbird/60.4.0
MIME-Version: 1.0
In-Reply-To: <2ddf4a32-517e-d6f4-d992-4a63b6035bfd@postgrespro.ru>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 11.02.2019 21:25, Arthur Zakirov wrote:
> Hello hackers,
> 
> Grigory noticed that one of our utilities has very slow performance when 
> xlogreader reads zlib archives. We found out that xlogreader sometimes 
> reads a WAL file block twice.
> 
> zlib has slow performance when you read an archive not in sequential 
> order. I think reading a block twice in same position isn't sequential, 
> because gzread() moves current position forward and next call gzseek() 
> to the same position moves it back.
> 
> It seems that the attached patch solves the issue. I think when reqLen 
> == state->readLen the requested block already is in the xlogreader's 
> buffer.
> 
> What do you think?
I looked at the history of the code changes:

---------------------------------------------------------------
7fcbf6a405f (Alvaro Herrera 2013-01-16 16:12:53 -0300 539) 
reqLen < state->readLen)

1bb2558046c (Heikki Linnakangas 2010-01-27 15:27:51 +0000 9349) 
        targetPageOff == readOff && targetRecOff < readLen)

eaef111396e (Tom Lane 2006-04-03 23:35:05 +0000 3842)
len = XLOG_BLCKSZ - RecPtr->xrecoff % XLOG_BLCKSZ;
4d14fe0048c (Tom Lane 2001-03-13 01:17:06 +0000 3843)
if (total_len > len) 
---------------------------------------------------------------

In the original code of Tom Lane, condition (total_len > len) caused a 
page reread from disk. As I understand it, this is equivalent to your 
proposal.
Th code line in commit 1bb2558046c seems tantamount to the corresponding 
line in commit 7fcbf6a405f but have another semantics: the targetPageOff 
value can't be more or equal XLOG_BLCKSZ, but the reqLen value can be. 
It may be a reason of appearance of possible mistake, introduced by 
commit 7fcbf6a405f.

-- 
Andrey Lepikhov
Postgres Professional
https://postgrespro.com
The Russian Postgres Company