XML CO527 Anonymous Questions and Answers Keyword Index

This page provides a keyword index to questions and answers. Clicking on a keyword will take you to a page containing all questions and answers for that keyword, grouped by year.

To submit a question, use the anonymous questions page. You may find the keyword index and/or top-level index useful for locating past questions and answers.

Keyword reference for multiprocessor

Question 16 (2010):

Submission reference: IN2020

Hi, I've been going through past papers & lecture slides and I have a few questions about multi-core/processor machines.

1. Would I be right in saying that since from the diagram on the slides Pentium processors appear to share their L2 cache, RAM accesses aren't cached local to a core so processes gain no performance from being scheduled to the same core as much as possible, or am I wrong in thinking that the cache is shared — perhaps it is the bus to the cache that is shared, with the cache divided up into sections for each core, just not stored on each core?
2. Similarly the slides say that AMD processors have to worry about page frame allocation more than Intel ones due to them having a non-uniform memory architecture. This seems to indicate that cores in AMD processors have fast, local RAM that only they can access — is this just the on core L2 cache? I always thought that the cache just, well, cached memory writes / accesses, so using some kind of virtual memory system for the cache seems a little odd, unless it just tries to maintain a local duplicate of relevant page frames in RAM? I'm probably fundamentally misunderstanding something here...

Thanks!

Answer 16:

1. A multicore package will probably share the L2 cache, but each core still has its own L1 cache. Of course, you might not be using a multicore machine (and just using regular multiprocessor), in which case there's L1 and L2 caches per core. As far as cache-affine scheduling in multi-processor multicore machines goes, it's probably wise to try and schedule the process on the same group of cores.
2. Whilst AMD (and now QPI based Intel machines as I mentioned in the lectures) have local and non-local RAM, any core can access any memory; it's just that non-local RAM costs more (in time) to access. From the diagram in lecture 5, non-local memory reads/writes come in over the hyper-transport bus (or QPI on modern Intels) and get routed to the memory, rather than up to the processor/core.

Keywords: multiprocessor