Mike,
thanks, I will do in this way. Just make sure I need to add this if.. else in every action, right?
Legion
On 2/20/07, Mike Marty <mikem@xxxxxxxxxxx> wrote:
I've added a directory cache to one of my own protocols. I added no states. Here is an example action. Notice all it does is check to see if dirCache contains the tag. If so, it uses DIRECTORY_CACHE_LATENCY instead
of DIRECTORY_LATENCY
action(f_forwardRequest, "f", desc="Forward request to owner") { peek(requestQueue_in, RequestMsg) { if (dirCache.isTagPresent(address) == false) {
enqueue(forwardNetwork_out, RequestMsg, latency="DIRECTORY_LATENCY") { out_msg.Address := address; out_msg.Type := in_msg.Type; out_msg.Requestor := in_msg.Requestor;
out_msg.Destination.addNetDest(directory[in_msg.Address].Owner); out_msg.AckCount := directory[address].Sharers.count(); if (directory[address].Sharers.isElement(map_L1CacheMachId_to_PrivateL2Cache(in_msg.Requestor)))
{ out_msg.AckCount := out_msg.AckCount - 1; } APPEND_TRANSITION_COMMENT(in_msg.Requestor); APPEND_TRANSITION_COMMENT(" "); APPEND_TRANSITION_COMMENT(directory[in_msg.Address].Owner);
out_msg.MessageSize := MessageSizeType:Forwarded_Control; } } else { enqueue(forwardNetwork_out, RequestMsg, latency="DIRECTORY_CACHE_LATENCY") { out_msg.Address := address;
out_msg.Type := in_msg.Type; out_msg.Requestor := in_msg.Requestor; out_msg.Destination.addNetDest(directory[in_msg.Address].Owner); out_msg.AckCount := directory[address].Sharers.count();
if (directory[address].Sharers.isElement(map_L1CacheMachId_to_PrivateL2Cache(in_msg.Requestor))) { out_msg.AckCount := out_msg.AckCount - 1; } APPEND_TRANSITION_COMMENT(in_msg.Requestor);
APPEND_TRANSITION_COMMENT(" "); APPEND_TRANSITION_COMMENT(directory[in_msg.Address].Owner); out_msg.MessageSize := MessageSizeType:Forwarded_Control; } }
} }
> Hi, > > I finally make the RAC work, at least the tester and splash2 didn't compain. > thanks for the helps. > Now I plan to add a directory cache. I do find one thread on this topic in
> the archive, and Mike recommended: > "You can change the getState() function in the Directory controller to return > a different state if the directory entry > is cached. Actions triggered by events to these cached states will use a
> lower latency". > > However, I have concerns to the above solution. First, we need to duplicate > all Actions/Transitions and make changes > everywhere. Second, maintaining a seperate set of states for directory cache
> is quite nontrivial, may even raise problems > at directory cache replacement. Therefore, I prefer to make changes in a > centralized place. > > Here are my random thoughts. In getState, if we make directory cache hit as
> the base case, and add some > latency (directory latency - directory cache hit latency) at misses, this > should be a close approximation. However, since > getState is not an event, I don't find a easy way to recycle (retry after xx
> cycles) it. As a result, I am thinking of adding > latency when poping the message from in-port. Basically, we add a new > function in MessageBuffer class say > cust-recycle(num_cycles). When peeking in-ports, we first check the
> directory cache. If tag matches, we pop the messages > as usual, else, we cust-recycle this message for (directory latency - > directory caches hit latency) cycles. Will this work? > > you guys are much more experienced on gems, any feedback/advice is much
> appreciated. > > thanks > Legion > _______________________________________________ Gems-users mailing list Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/
" to your search.
|