> I haven't solved the alteration issue yet. What was the roadblock to solving it the way I suggested originally? I'd be interested in why you cant delay the call to "setMemStall" in cache_unit.cc until the CompleteDataAccess command has been tried X times (with X being the cycles to the cache hit)
> I want to solve it by postponing > the tick event in the amount of the cache latency. That way when the > complete cache access stage will check the arrival of the cache data only > after the cache had time to respond to the request. I'm not sure I follow this. Do you want to postpone the tick event of the CPU? If you do that, then what about all the other instructions in the pipeline that need to process? What about the next thread you are switching to? Wouldnt it need the CPU's tick to process? It seems to me that if you want to delay this structurally, then what you do is add the necessary pipeline stages between your cache request and expectancy of the cache hit to be finished. > Now, what I haven't figured out yet, is how to access the cycle time and the > cache latency parameters. The ones I change through the python script. What I would do is fix the simple case first. Pick a latency you want and hardcode it. When you get that working, then work to get that passed down through the Python. > > Thanks, > Yuval. > > -----Original Message----- > From: [email protected] [mailto:[email protected]] On Behalf > Of Korey Sewell > Sent: Sunday, 22 July, 2012 20:45 > To: gem5 Developer List > Subject: Re: [gem5-dev] Inorder switch on cache hit > > Did you figure out your 1st problem w/altering the time for when a thread > switch will happen? I take it from your comments that you have a hardcoded > solution for that but you are trying to figure out a general solution to it. > > Is that you question or are you now elaborating on a different problem? From > your text, it's a little confusing on if you solved the 1st problem and are > working on a 2nd or are considering ways for the 1st problem. > > Lastly, I know some of the memory system code has been refactored (Andreas, > Ali?), but as far as I can remember the CacheUnit/CPU has a instruction and > data port attached to it. You can traverse one of those port until you get > to the Cache and then ask for the latency. > But again, I think the ports are changing so I need one of the memory system > guys to follow up on that. > > I'm not sure of a cleaner way to do that but you could also create your own > parameter to pass into the CPU, that waits X cycles before declaring a cache > miss. > > On Sun, Jul 22, 2012 at 10:29 AM, Yuval H. Nacson > <[email protected]> wrote: >> Hello all, >> >> >> >> I'm a relatively new user in gem5, but I already have my share of >> experience. >> >> I attached a conversation I had with Korey Sewell (which suggested me >> to move to this mailing-list) on the users list about the Switch on >> cache hit switching thread every memory access if the cache latency is >> longer than a cpu cycle. >> >> There's an example in the attached conversation. >> >> >> >> Now, What I want to do is to create a delay whenever a thread reaches >> for memory. For that I want to postpone the tick event of the cpu. >> Alas, I've yet to figure out how to reach the cache latency and CPU >> cycle time from within the code. >> >> >> >> Any suggestions or thought about my proposed solution? >> >> >> >> Thanks, >> >> Yuval >> >> >> >> >> >> Korey: >> >> Hi Yuval, >> >> Unfortunately, this isnt something you can switch in the config file >> (although it should be), so you'll need to edit some code to get this >> working. >> >> >> >> First, there is some code in pipeline_stage.cc that says asks an >> instruction if it's stalled on memory or not. I'm guessing >> "inst->isMemStall()". This function needs not to return true after only 1 > cycle. >> >> >> >> Second, that function is set to true in cache_unit.cc. In the execute >> function, you'll probably see a "inst->setMemStall()" in there if it >> tries to complete the access and the data isnt available yet. >> >> >> >> What you'll need to do is edit that code in cache_unit.cc to just wait >> the necessary amount of cycles before declaring a memory stall. Please >> move the conversation to gem5-dev if you want to go the route of >> submitting the patch to make the "MemStall" general instead of 1 cycle. >> >> >> >> On Mon, Jul 16, 2012 at 10:07 AM, Yuval H. Nacson >> <[email protected]> wrote: >> >>> Hey, >> >>> >> >>> >> >>> >> >>> In my quest of inorder pipeline I've encountered another problem. >> >>> >> >>> I run two threads on a switch on cache miss machine. Also. as opposed >> >>> to other question I've sent so far, I'm using the 5 stages pipeline >> >> supplied >> >>> with the simulator. >> >>> >> >>> The Frequency is 2Ghz and I set the cache latency to be 800ps (which >> >>> is >> >> more >> >>> than a clock cycle). >> >>> >> >>> >> >>> >> >>> Now. In the fourth stage memory request is being sent and in the >>> fifth >> >> stage >> >>> the cache is being read. >> >>> >> >>> I get the following: >> >>> >> >>> >> >>> >> >>> 159888500: system.cpu.stage4: [tid:0]: Not blocked, so attempting to >> >>> run stage. >> >>> >> >>> 159888500: system.cpu.stage4: [tid:0]: Processing instruction >> >>> [sn:41180] >> >> ldq >> >>> with PC (0x1200003f0=>0x1200003f4) >> >>> >> >>> 159888500: system.cpu.stage4: [tid:0]: [sn:41180]: sending request to >> >>> system.cpu.dcache_port. >> >>> >> >>> 159888500: system.cpu.dcache_port: [tid:0]: [sn:41180]: Updating the >> >> command >> >>> for this instruction >> >>> >> >>> 159888500: system.cpu.dcache_port: [tid:0]: [sn:41180]: Trying to >> >> Complete >> >>> Data Read Access >> >>> >> >>> 159888500: system.cpu.dcache_port: STALL: [tid:0]: Data miss from >> >>> 0x12009ed78 >> >>> >> >>> 159888500: system.cpu.stage4: [tid:0]: [sn:41180] request to >> >>> system.cpu.dcache_port failed. >> >>> >> >>> 159888500: system.cpu.stage4: [tid:0] [sn:41180] Detected cache miss. >> >>> >> >>> 159888500: system.cpu.stage4: Inserting [tid:0][sn:41180] into switch >> >> out >> >>> buffer. >> >>> >> >>> 159888500: system.cpu: Scheduling CPU Event (SquashFromMemStall) for >> >> cycle >> >>> 159888500, [tid:0]. >> >>> >> >>> 159888500: system.cpu.ResourcePool: Ignoring Unrecognized CPU Event >> >>> (SquashFromMemStall). >> >>> >> >>> 159888500: system.cpu.stage4: Suspending [tid:0] due to cache miss. >> >>> >> >>> >> >>> >> >>> . >> >>> >> >>> >> >>> >> >>> 159888800: system.cpu.dcache_port: [tid:0]: [sn:41180]: [slot:4] >> >>> Waking >> >> from >> >>> cache access (vaddr.0x12009ed78, paddr:0x090d78) >> >>> >> >>> 159888800: system.cpu.dcache_port: [tid:0]: [sn:41180]: Processing >> >>> cache access >> >>> >> >>> 159888800: system.cpu.dcache_port: [tid:0]: [sn:41180]: Bytes loaded >> >> were: >> >>> 0000000000000000 >> >>> >> >>> 159888800: system.cpu.dcache_port: [tid:0] Waking up from Cache Miss. >> >>> >> >>> 159888800: system.cpu: [tid:0]: Activating ... >> >>> >> >>> >> >>> >> >>> Note the Time differences. >> >>> >> >>> I would like to have my cache to have a latency of 2-3 cycles and not >> >> having >> >>> the thread switched. Can it be done? >> >>> >> >>> >> >>> >> >>> Thanks, >> >>> >> >>> Yuval >> >>> >> >>> >> >>> _______________________________________________ >> >>> gem5-users mailing list >> >>> [email protected] >> >>> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >> >> >> >> >> >> >> >> -- >> >> - Korey >> >> _______________________________________________ >> >> gem5-users mailing list >> >> [email protected] >> >> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >> >> >> >> _______________________________________________ >> gem5-dev mailing list >> [email protected] >> http://m5sim.org/mailman/listinfo/gem5-dev > > > > -- > - Korey > _______________________________________________ > gem5-dev mailing list > [email protected] > http://m5sim.org/mailman/listinfo/gem5-dev > > _______________________________________________ > gem5-dev mailing list > [email protected] > http://m5sim.org/mailman/listinfo/gem5-dev -- - Korey _______________________________________________ gem5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/gem5-dev
