[bgl-discuss] RE: [bgl-support #151] buffering

[Ray Loy <rloy@xxxxxxxxxxx>]

> I am not sure I asked this already, but are the reads/writes buffered
> on the Blue Gene ?


waris,

no, unless you count going through the IO node (see copy below).
maybe someone on bgl-discuss will have more details.


ray



> To: "Sindhi, Abdul Waris" <abdul.sindhi@xxxxxxxxxx>
> cc: BG/L Support <support@xxxxxxxxxxxxxxx>, acherry@xxxxxxxxxxx,
>     smc@xxxxxxxxxxx
> Subject: Re: [bgl-support #151] Slow I/O on the blue gene 
> In-reply-to: <7084ED0DD428AF4DB78B7B7E88A732910E326646@xxxxxxxxxxxxxxxxxxxx> 
> References: <7084ED0DD428AF4DB78B7B7E88A732910E326646@xxxxxxxxxxxxxxxxxxxx>
> Comments: In-reply-to "Sindhi, Abdul Waris" <abdul.sindhi@xxxxxxxxxx>
>    message dated "Thu, 30 Nov 2006 13:02:44 -0500."
> Date: Thu, 30 Nov 2006 12:14:40 -0600
> From: Ray Loy <rloy@xxxxxxxxxxx >
> 
> 
> > By 'each slave' I meant each of the processors in the job.
> > 
> > Do the read/write requests from the compute nodes that go 
> > to the I/O node get buffered ? 
> > 
> > Waris
> 
> 
> The low-level read() and write() calls on the compute-node actually
> transfer the data to the IO node via the tree network, and from there
> it goes to the disk server.  I don't think there is any more buffering
> at the IO node than strictly necessary to complete the operation.
> 
> I agree with Andrew's suggestion to try a /pvfs destination directory.
> 
> 
> Ray
> 
> 

- --------------------------------------------------------------------
To add or remove yourself from this mailing list, use the 'notifyme'
command on any BGL machine. To remove: notifyme -n, to add: notifyme -y.