Do modern multiprocessors only support a fixed number of outstanding requests as our processor did (and if so, what's the number?), or do they work with a dynamic number of requests?
This comment was marked helpful 0 times.
There will be a fixed number of outstanding requests. Hardware bookkeeping structures that keep track of outstanding requests (e.g., the request table in this lecture) will be fixed size.
Take a look at this figure for some stats from a recent Core i7. The number of outstanding misses in the figure is the number of outstanding memory requests from any one cache. This is slightly different from the number of outstanding requests on the interconnect, which we deal with in this lecture, but it makes a similar point.
It seems like we don't really talk about the issue of possible starvation due to the NACK. I would guess that in real hardware there are procedures in place to ensure that you don't get a NACK an unbounded number of times.
So the system queues request regardless of request content? If so the chance of starvation can probably be kept low just by the design of hardware.
A simple solution is that you elevate the priority of a client each time they get NACKed. (or decrease the priority of a client each time they successfully use the bus) If the bus arbitrator inspects these priorities when making bus grant decisions, no one will be starved.
I might be misunderstanding something here, but don't you get a NACK if the buffer is full? The arbitration is then solely on whether the buffer is full or not, so I'm not understanding how increasing priority would work - if you're unlucky and issue a request just as the buffer fills up every time, you'll just get more NACKs.
Perhaps you could simply inspect the entire buffer and throw out requests that have lower priority than the incoming one, but this could lead to livelock if the thrown out request causes many more requests that get NACK'd and its priority eventually increasing higher than the request that had originally evicted it from the buffer.
edit: on second thought, it seems like decreasing the priority of the client on each successful bus use would work pretty well in combination with inspection of the buffer. Nevermind.