From jdonald at Princeton.EDU Mon Dec 8 17:02:41 2003 From: jdonald at Princeton.EDU (James Donald) Date: Wed Mar 22 16:04:09 2006 Subject: [Hotspot] Plugging HotSpot into Wattch 1.02 Message-ID: <46f48446f07f.46f07f46f484@Princeton.EDU> I'm attempting to get HotSpot running with Sim-Wattch so I can get some sanity-check numbers out with the given floorplan, but it seems there is initially quite a bit of hacking that I need to do. Perhaps due to the enhancements made to Sim-Wattch as described on page 22 of the HotSpot technical report, the processor's building blocks in the HotSpot template code do not match very nicely with power number variables as found in the version of Sim-Wattch Wattch that I am using. From ks4kk at cs.virginia.edu Mon Dec 8 18:36:22 2003 From: ks4kk at cs.virginia.edu (Karthik Sankaranarayanan) Date: Wed Mar 22 16:04:09 2006 Subject: [Hotspot] Plugging HotSpot into Wattch 1.02 In-Reply-To: <46f48446f07f.46f07f46f484@Princeton.EDU> References: <46f48446f07f.46f07f46f484@Princeton.EDU> Message-ID: > But plenty of them aren't so simple matchings. My first question is, how > is the L2 cache divided into three functional units? What functions are > performed by "left", "right", and "bottom"? Sim-Wattch's default > configuration has only power.dcache2_power. We took the L2 power given by Wattch and computed the power density based on the total L2 area from the Alpha-like floorplan. To come up with the power numbers for the individual L2, chunks, we just assume uniform power density and hence distribute the total power weighted by the areas of the inidividual chunks. > To get started here and make sure I can get some realistic results with > HotSpot, I'm inclined to edit the floorplan and combine pieces such as > IntReg+FPReg or FPAdd+FPMul back into monolithic units. Certainly, to > accurately model modern processors, later on I should make sure to later > on simulate real issues queues and more separated structures. But just > to get started for this week, would it make sense to hack things > together (no real issues queues, and some separate structures made > monolithic) and still hope to get some reasonable sanity check numbers > out? Yes - that is totally reasonable. > Lastly, how should I get power numbers for ITB and DTB, which I assume > are the Integer TLB and Data TLB? Yes - they are the I and D TLBs. > And what are "IntMap" and "FPMap"? They are the Integer and FP register renamers. Hope this helps. -karthik