Bernard Leclerc (bleclerc++at++cae.ca)
Tue, 8 Oct 1996 11:30:35 -0400
> I have a very large structure of shared memory used between processors
> since I share a great deal of data between application and draw
> processes running on separate processors. As I have just been tasked
> to add some more data that will need to be shared between application
> processes and draw processes, I have begun to worry about just how much
> data I'm sharing and what overhead is associated with it.... I just
> passed 240K of shared data (allocated with pfMalloc) and I am heading
> quickly towards 300K. What is my transfer overhead between CPU's with
> this shared structure?
>
> I am using an Infinite Reality with 16 R10000 CPUs and Performer 2.1.
> My application uses a minimum of 5 app processes on different CPUs (only
> one does model placement and other geometry operations -- 4+ handle
> higher math for calculating model placement) with dual pipe (which I
> gather is handled with one draw process per pipe?). The 5 app processes
> handle 60% of the data between each other and the draw processes
transfer
> the remaining to the app process. Right now, this is one allocation
> transferring the bulk of EVERYTHING to ALL processes which is highly
> inefficient, but out of lack of time, I have never changed it. Should I
> worry about transferring ~300K of data around 8 to 16 processors?
Jeff,
If you're using shared memory between your processes, you don't need to
worry since the data is not physically copied from producers to consumers.
In fact, this is probably the best approach in sharing a large amount of
data between related processes.
--
Bernard Leclerc CAE Electronics Ltd., 8585 Cote De Liesse
Technical Leader Saint-Laurent, Quebec, Canada, H4L-4X4
3-D Graphics Applications tel: +1 514 341 2000 extension 2275
bleclerc++at++cae.ca fax: +1 514 340 5496
=======================================================================
List Archives, FAQ, FTP: http://www.sgi.com/Technology/Performer/
Submissions: info-performer++at++sgi.com
Admin. requests: info-performer-request++at++sgi.com
This archive was generated by hypermail 2.0b2 on Mon Aug 10 1998 - 17:53:43 PDT