From: Jim Starkey
Date: January 28 2009 7:38pm
Subject: Re: Problems with record visibility and how it is computed
List-Archive: http://lists.mysql.com/falcon/440
Message-Id: <4980B42F.5070409@nimbusdb.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Olav Sandstaa wrote:
>>
>> There isn't much difference in cost between a lock(Shared)/unlock and 
>> addRef/release.  Each requires a pair of interlocked instructions.
>
> Thanks for the clearification about the cost. I expected the interlock 
> instruction to be relatively cheaper than a shared lock.
>
>
My performance testing suggests that on Intel and AMD, an interlocked 
CAS is cheaper than an interlocked increment or decrement.  I suspect 
this is because the interlocked CAS is a single memory operation that 
either succeeds or fails in the memory controller while the interlocked 
increment (actually a "add 1 to memory") requires a memory bus lock 
while the processors fetches the operand, does the add, and writes it back.

This further suggests that we might get some bang by replacing our 
interlocked increment and decrement with interlocked CASes.


-- 
Jim Starkey
President, NimbusDB, Inc.
978 526-1376