Here you can ask questions and find or give answers to organizational, academic and other questions about studying computer science.

1.1k questions

1.2k answers

1.6k comments

531 users

0 votes
On slide 13 of the Dynamic Scheduling chapter it is spoken about different and specialized processing unit (in the following PU). Now assume we have PU1 and PU2 where PU1 is able to do all kinds of operations but requires many cycles to complete (for example 9 cycles), while PU2 can only do addition but only needs for example 2 cycles.

Now, think of a sequence of two addition instructions. The first instruction obviously makes use of PU2 and is finished after 2 cycles.
The question is now, if the second instruction waits for the first to be finished in order to also use the fast PU2 or if it is a "greedy system" so that the second instruction immediately uses PU1 so that it can start earlier but finished much later.
Assuming a third instruction which needs for example multiplication follows, the consequences would be even more extreme, since this instruction would have to wait for the second instruction to finish and make PU1 available.

Thank you for your answer!

Kind regards
in * TF "Emb. Sys. and Rob." by (150 points)

1 Answer

+1 vote

This is an excellent question!

However, on second thought, this does not make that much sense to me, since the size of the function units is not so problematic anymore. So if we can implement addition in 2 cycles, we would use that implementation also in PU1, where addition will then also take 2 cycles, even though other operations may take longer. If the PUs were implemented like this, it would not matter to which PU we would send the addition (at least concerning the completion time of the two additions). However, PU1 would be blocked for other instructions, so we might regret using PU1 for the additions when other instructions would follow.

All in all, I am not aware of any optimized dispatching that makes use of such considerations. Dispatching units also need to be kept as simple as possible, so I think they just send operations to the next free unit they find. Many processors also use distributed reservation stations where each PU has an input queue of operands. Then, dispatching is even done earlier and can just consider how many operations are queueing up.

Also, I am not aware of any PUs with "overlapping functionality", i.e. that one PU can do what another can also do, but also other operations as well. Instead, we usually find ADD/SUB/COMPARE on some units, and MULT and DIV on others. So, then, the problem does not appear. 

Anyway, it is an interesting question that could be thought over to further optimize dynamic scheduling! However, since we do not foresee the next instructions, we may always regret the scheduling done so far, so I doubt that there is an optimal solution for online scheduling (which is what we would have to do).

See also 

by (166k points)
Imprint | Privacy Policy
...