Performance on NUMA

----- Original Message -----
From: "Luis Alberto Amigo Navarro" <lamigo(at)atc(dot)unican(dot)es>
To: "Neil Conway" <neilc(at)samurai(dot)com>
Cc: <pgsql-hackers(at)postgresql(dot)org>
Sent: Wednesday, November 06, 2002 5:02 PM
Subject: Re: [HACKERS] a tiny question

> A small patch for low performance on NUMA arquitectures could be the
chance
> of using more than one shared region.
> Several months away there was a brief talk about low performance on IRIX,
it
> was not real, it's low performance on Origin servers, they use ccNUMA
> architecture, so using a single shared segment allocated by only one
> proccess doesn't give IRIX planner a chance for attempting to optimize
> memory accesses, even thought page migration is almost impossible with DB
> memory accesses.
> With this scenario, IRIX uses a first-touch algorith for memory placement
> (all the memory is allocated on the node Postgres starts running). Due to
> high latency between local and remote memory accesses, if there is data
> modification, remote proccesors cache hit rate goes down, all of this
makes
> that all the basis of high scalability on this systems(big cache, memory
> distribution, increasing bandwith and page migration) ar thrown to the
> earth.
> I suppose that making more than one shared segment it's not too difficult,
> but with this u can do that a long time Postgres running on a NUMA system
> achieve a good performance scenario.
> Is it so difficult?
> Thanks and regards.
> Luis Amigo
>
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 2: you can get off all lists at once with the unregister command
> (send "unregister YourEmailAddressHere" to majordomo(at)postgresql(dot)org)
>
>