In class someone mentioned the idea using address as the tag, instead of using the entry of the table. I still think it is a good idea, since we already have the address bus at the request bus. So what is the disadvantage of such method?
This comment was marked helpful 0 times.
mchoquet
Maybe it's just that the response tag is smaller? It looks like like lookup in the request table is done by-tag, so it might be simpler to choose a smaller tag than the size of a memory address.
This comment was marked helpful 0 times.
yrkumar
@yetianx: At first, I thought that problems might occur with that approach when multiple outstanding requests existed in the request table for a single address (i.e. different processors request to read the same address). However, I now realize that if there are multiple requests to the same address, then the responses are the same for both of those requests, meaning that they don't need to be matched up to the specific processor that made the request and could just be sent to whichever processors have outstanding requests to that address. This brings us back to the same question. It seems redundant to have an extra bus line for the response tag when the address could serve as the identifier.
This comment was marked helpful 0 times.
kleino
Well I think the point is that since we also have to keep track of the outstanding requests, and therefore will have an index into the table of requests, we may as well use the index because it's smaller. Basically, there is no advantage to using the address over the tag, so we may as well use fewer wires for the response bus. Note that the address is not part of the response bus already - the address is only included in the request bus - so we don't get the address for free.
This comment was marked helpful 0 times.
yetianx
@kleino: I think your point that we don't get the address for free is right. If we use the address line to pass the response memory address, we lost the ability to pass the request memory address at the same time. It would slow our pipeline.
Now, another process can send a request while one process is receiving data.
This comment was marked helpful 0 times.
ycp
On modern CPU's, how many buses are there? (Like it says on the slide the BUS is split into a Request and Response BUS, and then there is also the response tag bus).
@yetianx: In hardware selecting an entry from that table using its index needs only a simple multiplexer. Doing this using addresses, would require a comparator for each entry in the table (to check for address matches). This would be more costly in silicon, and time.
In class someone mentioned the idea using address as the tag, instead of using the entry of the table. I still think it is a good idea, since we already have the address bus at the request bus. So what is the disadvantage of such method?
This comment was marked helpful 0 times.
Maybe it's just that the response tag is smaller? It looks like like lookup in the request table is done by-tag, so it might be simpler to choose a smaller tag than the size of a memory address.
This comment was marked helpful 0 times.
@yetianx: At first, I thought that problems might occur with that approach when multiple outstanding requests existed in the request table for a single address (i.e. different processors request to read the same address). However, I now realize that if there are multiple requests to the same address, then the responses are the same for both of those requests, meaning that they don't need to be matched up to the specific processor that made the request and could just be sent to whichever processors have outstanding requests to that address. This brings us back to the same question. It seems redundant to have an extra bus line for the response tag when the address could serve as the identifier.
This comment was marked helpful 0 times.
Well I think the point is that since we also have to keep track of the outstanding requests, and therefore will have an index into the table of requests, we may as well use the index because it's smaller. Basically, there is no advantage to using the address over the tag, so we may as well use fewer wires for the response bus. Note that the address is not part of the response bus already - the address is only included in the request bus - so we don't get the address for free.
This comment was marked helpful 0 times.
@kleino: I think your point that we don't get the address for free is right. If we use the address line to pass the response memory address, we lost the ability to pass the request memory address at the same time. It would slow our pipeline.
Now, another process can send a request while one process is receiving data.
This comment was marked helpful 0 times.
On modern CPU's, how many buses are there? (Like it says on the slide the BUS is split into a Request and Response BUS, and then there is also the response tag bus).
This comment was marked helpful 0 times.
Woops, missed this, Intel's Ring Interconnect has 4 rings.
This comment was marked helpful 0 times.
@yetianx: In hardware selecting an entry from that table using its index needs only a simple multiplexer. Doing this using addresses, would require a comparator for each entry in the table (to check for address matches). This would be more costly in silicon, and time.
This comment was marked helpful 2 times.