From: Kevin Lewis
Date: December 15 2008 4:53pm
Subject: Re: Falcon and ZFS
List-Archive: http://lists.mysql.com/falcon/309
Message-Id: <49468B8B.4@sun.com>
MIME-Version: 1.0
Content-Type: text/plain; format=flowed; charset=ISO-8859-1
Content-Transfer-Encoding: 7BIT

Jim Starkey wrote:
> If Falcon were to allocate blocks of pages (say, 64K's worth) per data 
> segmentmd per index and maybe section tree, we would get greater 
> locality on very large page size file systems like ZFS, and probably not 
> hurt us at all other systems.  It wouldn't have any affect on 99% of 
> Falcon, and would probably be easy to prototype.

Jim,

I think this idea has been captured previously by

WL#4564 - Falcon; Page-Allocation by groups.

Let me know if you think this high level description needs to be 
expanded or changed;

  High-Level Description

It is generally accepted that that reading 64K is effectively just as 
fast as reading a 4K page.  Falcon contains an optional 'Sector cache' 
which tries to take advantage of that fact.  This was added to version 
6.0.5 in June, 2008.  But in subsequent tests, we hardly ever found a 
performance benefit.  In practical application, the overall cost of an 
extra cache in front of the page cache was worse than the benefit of 
having already read a few more pages than needed, allowing the engine to 
skip another read every now and then.

The page allocation scheme within a tablespace needs to be improved in 
order to benefit from this sector cache.  We need to increase the 
likelihood that consecutive pages on disk contain common data.   The 
current scheme allocates one page at a time from the tablespace to a 
particular table.  In a busy system, pages assigned to one table are 
interleaved with pages assigned to completely different tables.  Falcon 
should assign some configurable number of pages to a table at a time. 
This will greatly increase the chance that new pages with common or 
dependent data are located closer together on disk.  And it would
improve the possible payback of using the current 'sector cache' scheme 
of read-ahead.