HotSpot 4.0 Changes ------------------- This version of HotSpot incorporates several accuracy enhancements described in the following paper: W. Huang, K. Sankaranarayanan, R. J. Ribando, M. R. Stan and K. Skadron. "An Improved Block-Based Thermal Model in HotSpot 4.0 with Granularity Considerations." In Proceedings of the Workshop on Duplicating, Deconstructing, and Debunking, in conjunction with the 34th International Symposium on Computer Architecture (ISCA), June 2007. While the paper itself focuses on the block model, some of the changes also improve the accuracy of the grid model. Specifically, following is a list of the modifications: a) Accuracy Enhancements: 1) Enhanced package modeling: removal of the forced isotherm 2) Modeling of lateral heat flow in the TIM layer 3) Block model: a wrapper script to sub-divide floorplan blocks with high aspect ratio to improve accuracy 4) Grid model: upgradation of the first order transient sover to a fourth order adaptive step-size Runge-Kutta solver 5) Grid model: an option to maintain grid thermal state across calls to the transient solver b) Updated defaults: The default package parameters have been updated to reflect contemporary packages. The most significant of the changes are: 1) Die thickness has been updated to 0.15 mm down from 0.5 mm. (See: B. Majeed, I. Paul, K. M. Razeeb, J. Barton and S. C. O'Mathuna. "Microstructural, Mechanical, Fractural and Electrical Characterization of Thinned and Singulated Silicon Test Die". Journal of Micromechanics and Microengineering, 16:1519-1529, Aug. 2006.) 2) Tim thickness and resistivity have been updated respectively to 20u down from 75u and 0.25 mK/W down from 0.75 mK/W. (See: E.C. Samson, S.V. Machiroutu, J.-Y. Chang, I. Santos, J. Hermerding, A. Dani, R. Prasher, and D.W.Song. "Interface Material Selection and a Thermal Management Technique in Second-Generation Platforms Built on Intel Centrino Mobile Technology." Intel Technology Journal, 9(1), Feb. 2005.) c) Others: 1) Complete re-write of the grid model for easy readability and extensibility. 2) Minor bugfixes etc. Installation ------------ 1. Download the HotSpot tar ball (HotSpot-4.0.tar.gz) following the instructions at http://lava.cs.virginia.edu/HotSpot 2. Unzip and untar the file using the following commands a) gunzip HotSpot-4.0.tar.gz b) tar -xvf HotSpot-4.0.tar 3. Go to the HotSpot installation directory a) cd HotSpot-4.0/ 4. Uncomment the lines corresponding to your installation of the BLAS/LAPACK vendor libraries (see http://www.netlib.org/blas/faq.html#5) and set the path and compiler options corresponding to your library. This version of HotSpot has code integrated for Intel Math Kernel Library, AMD Core Math Library, Apple Velocity Engine and Sun Performance Library. Extending it to other vendors should be straightforward. For such an extension, apart from the user guide from those vendors, http://www.netlib.org/blas/blast-forum/cinterface.pdf might also be useful as it provides useful description of the C interface to BLAS. 5. Build HotSpot a) make 7. To remove all the outputs of compilation, type 'make clean'. To remove the object files alone, type 'make cleano'. To view the list of files HotSpot needs for proper working, type 'make filelist'. To compile for debugging, use 'make DEBUG=1'. To compile for using with a profiler (e.g. gprof), use 'make DEBUG=2'. 8. Known compatibility issues: a) For old AMD machines without SSE2 instructions, the most recent version of ACML available is 3.1.0. On such machines, the ACML library works with GCC 4.0 but not with GCC 4.1. b) Linking with Sun Performance Library on old Solaris machines fails as 'libmtsk' was not found. (e.g. see http://supportforum.sun.com/jive/thread.jspa?threadID=72529) Installing patch 117560 from Sun's patch finder (http://sunsolve.sun.com/pub-cgi/show.pl?target=patches/patch-access) solves the problem. Notes on the Modifications: --------------------------- 1) As the WDDD paper above mentions, the block model suffers from spatial discretization errors when the modeled blocks have very high aspect ratios. In such cases, the user should either use the grid model or the floorplan should be modified in such a manner that the blocks are sub-divided into square-shaped sub-blocks. To aid in this aspect ratio sub-division, this version provides two wrapper scripts: hsconvert-flp.pl and hsconvert-data.pl. The former can be used to convert floorplans with high-aspect-ratio blocks while the latter can be used to migrate power/temperature trace data corresponding to the old floorplan into the new floorplan. Here is a sample use where ev6.flp.orig is the old floorplan and the power trace corresponding to it is gcc.ptrace.orig. The following sequence of commands produce a floorplan file (ev6.flp) with nearly square-shaped blocks and a compatible power trace file (gcc.ptrace): hsconvert-flp.pl ev6.flp.orig > ev6.flp hsconvert-data.pl -o ev6.flp.orig -c ev6.flp -p gcc.ptrace.orig > gcc.ptrace It is to be noted that being a finite-element solver, HotSpot cannot avoid discretization errors altogether. However, they can be dealt with effectively as a speed vs. accuracy trade-off. For instance, in cases where higher accuracy is desired, the grid model could be used. It is to be noted that even with the grid model, high aspect ratios can be a problem. A common pitfall is using an extremely fine mesh in the x-y direction without correspondingly increasing the resolution in the z direction. For instance, if one must use a 1000 x 1000 grid, then the no. of z-layers should also be increased using the grid layer file specification so that each volumetric cell is still nearly "cubical" and not "tall and skinny". 2) The interfaces to HotSpot remain almost identical to previous versions. The minor differences include: a) The grid layer file specification has been made optional with the default being one silicon layer and one TIM layer (similar to the block model). This change also means that the grid model is now compatible with HotFloorplan. However, this compatibility is only theoretical because, for practical grid sizes, the grid model would be too slow to be used with HotFloorplan. b) Before solving for transient temperature, the grid model maps the functional block power and temperature numbers onto each grid cell. After the solution is done, the temperature numbers are mapped back onto each functional block. This reverse mapping involves averaging. When the transient solver is called repeatedly (as is usually the case), and if the internal state of the solver is not maintained across calls, this averaging can result in some loss of accuracy. To avoid that, this version of HotSpot provides an option to maintain state across calls to the transient solver of the grid model. When the solver is called for the first time, the functional block temperature vector is passed as an argument to compute_temp. During subsequent calls however, a NULL pointer is passed signalling the solver to maintain its internal state from the last call and continue passing on its output to the same temperature vector provided earlier. Here is the relevant code snippet from hotspot.c that does this: /* first time */ if (model->type == BLOCK_MODEL || lines == 0) compute_temp(model, power, temp, model->config->sampling_intvl); /* subsequent calls - 'temp' still gets updated every time with the * transient solution */ else compute_temp(model, power, NULL, model->config->sampling_intvl);
