From nishchay.bh at gmail.com Mon May 7 18:35:18 2007 From: nishchay.bh at gmail.com (Nishchay Bharati) Date: Mon May 7 18:35:24 2007 Subject: [Hotspot] HOTSPOT with benchmarks Message-ID: <507e82580705071535q76a95731q41be355738b91b22@mail.gmail.com> Hi Everyone, I have just started using HOTSPOT and I wanted to know how I can run HOTSPOT with SPEC2006 benchmarks. I mean how should I modify the command line for HOTSPOT to see how the different units of the processor heat up to different benchmarks. Thanks a lot, Nishchay -------------- next part -------------- An HTML attachment was scrubbed... URL: http://www.cs.Virginia.EDU/pipermail/hotspot/attachments/20070507/9401aea3/attachment.html From guptaaseem at gmail.com Mon May 7 18:43:32 2007 From: guptaaseem at gmail.com (Aseem Gupta) Date: Mon May 7 18:43:44 2007 Subject: [Hotspot] HOTSPOT with benchmarks In-Reply-To: <507e82580705071535q76a95731q41be355738b91b22@mail.gmail.com> References: <507e82580705071535q76a95731q41be355738b91b22@mail.gmail.com> Message-ID: i guess u can use PTScalar -aseem On 5/7/07, Nishchay Bharati wrote: > Hi Everyone, > I have just started using HOTSPOT and I wanted to know > how I can run HOTSPOT with SPEC2006 benchmarks. I mean how should I modify > the command line for HOTSPOT to see how the different units of the processor > heat up to different benchmarks. > > Thanks a lot, > Nishchay > > _______________________________________________ > HotSpot mailing list > HotSpot@ares.cs.Virginia.EDU > http://www.cs.Virginia.EDU/mailman-2.1.5/listinfo/hotspot > > From wh6p at cms.mail.virginia.edu Tue May 8 15:06:28 2007 From: wh6p at cms.mail.virginia.edu (Wei Huang) Date: Tue May 8 15:06:35 2007 Subject: [Hotspot] Re: Hotspot Query In-Reply-To: References: <457828DC.6090006@mail.cs.virginia.edu> <592D759C9E9D96196E8D2117@mstu2> <96F600E1A1994187CE305A89@mstu2> <511C87CF989B30C59017CE8B@mstu2> <3967F151E6930F56B8CABB78@mstu2> <8707E5418AEC52D23FD00E96@mstu2> <3B4C9497E4863D4C01D2BC31@mstu2> <9B6257EE5E9AFE20B844A4A1@mstu2> Message-ID: Hi Niti, glad we can help! Please see my replies below. (Karthik may have more insights.) -Wei > 1) Typically, what is the recommended temperature metric > to report in papers - steady-state temperatures or > the average of transient temperature? I have seen > some papers where the authors advocate running Hotspot > twice after initiliazing the second run with first run's > steady-state temperatures. But they report the second > run's steady-state temperature. Shouldn't the > steady-state temperatures be the same in both runs? I > see only transient temperatures to be different in 2 runs. I can't see why 2 runs are needed for steady temperatures. Can you point me to the papers you mentioned? Normally, steady-state temperatures are useful for initializing transient thermal simulations. But you still need some additional "warm-up" simulations after setting initial temperatures to steady-state ones. Steady-state temperature is also useful when transient power numbers are not available or too expensive to obtain. I guess this is not your case, but for circuit/SOC/ASIC designers, this is usually true. Because for them, doing clock-cycle-accurate circuit simulations to model transient temperature needs a super powerful and detailed transient power estimator and super long simulation time. > > 2) For a 5 us sampling interval, the transient > temperature for hotter blocks is higher than the > steady-state temperature in the second run. I saw a few > mails in your online archived mailing list about > steady-state temperature not reaching. According to > that, a much larger sampling interval is recommended for > transient and steady-state results to match. Is there any > recommended value for this interval? In HotSpot, we use average power to calculate steady temperatures. So it is normal to see transient temperatures higher or lower than steady ones. Using a large sampling interval is essentially equivalent to averaging the power trace over a longer time, so it is not surprising to see that this matches the steady temperature better. By and large, you can trust your transient simulations (provided with a long enough "warm-up" stage). Steady temperature is more of a "reference" that can be useful in some cases as I mentioned before. > > 3) I wanted to model DTM in my implementation. I need to > use transient temperatures to trigger the DTM in the > second run. I saw a paper in WDDD: > > http://www.ece.wisc.edu/~wddd/2005/papers/WDDD05_srinivas > an.pdf According to this paper, their tool as well as > Hotspot doesn't model reactive DTM if used with only 2 > runs as the heat-sink model is not initialized > adaptively for DTM. They recommend 5-6 runs to model > reactive DTM. Do you agree with this methodology? Again, I think this is related to the "warm-up". The warm-up stage has to be long enough to "heat up" the heatsink. So you may want to monitor the heat sink temperature to see if it stabilizes (i.e. some of the "EXTRA" nodes in HotSpot code). However, you may also try to tweak the thermal capacitance of the heatsink to small values during the warm-up phase, so that it artificially brings the heat sink to a steady temperature much faster. But you need to be really careful with this since it is error-prone. I don't recommend this unless you are really familiar with the HotSpot code base and how it works. > > 4) Finally, the DTM macro in Hotspot, does it model > heat-sink initialization adaptively if enabled? I have > never used it and wasn't sure how it works. I am not 100% percent sure. Karthik should be more helpful than I am on this issue. But I assume this is the case since the model itself already has all the structures needed for this. Hope this helps. > > Thanks for all your help. > > -Niti > From ks4kk at cs.virginia.edu Tue May 8 15:22:48 2007 From: ks4kk at cs.virginia.edu (Karthik Sankaranarayanan) Date: Tue May 8 15:22:52 2007 Subject: [Hotspot] Re: Hotspot Query In-Reply-To: References: <457828DC.6090006@mail.cs.virginia.edu> <592D759C9E9D96196E8D2117@mstu2> <96F600E1A1994187CE305A89@mstu2> <511C87CF989B30C59017CE8B@mstu2> <3967F151E6930F56B8CABB78@mstu2> <8707E5418AEC52D23FD00E96@mstu2> <3B4C9497E4863D4C01D2BC31@mstu2> <9B6257EE5E9AFE20B844A4A1@mstu2> Message-ID: >> 1) Typically, what is the recommended temperature metric >> to report in papers - steady-state temperatures or >> the average of transient temperature? I have seen >> some papers where the authors advocate running Hotspot >> twice after initiliazing the second run with first run's >> steady-state temperatures. But they report the second >> run's steady-state temperature. Shouldn't the >> steady-state temperatures be the same in both runs? I >> see only transient temperatures to be different in 2 runs. I am not sure which paper you are referring to but if the leakage-temperature feedback loop is modeled, then this can happen. Since leakage is dependent on temperature and the next iteration's steady state temperature is dependent on the leakage, one has to weight for convergence. >> 4) Finally, the DTM macro in Hotspot, does it model >> heat-sink initialization adaptively if enabled? I have >> never used it and wasn't sure how it works. I am not sure what you mean by adaptive here. In HotSpot, the dtm_used flag enables clipping while reading in temperatures. This means that if the steady state temperatures are above the thermal threshold, they (all of them - including heat sink temperatures) are linearly scaled down so that the maximum temperature on the die is now the thermal threshold itself. Hope this helps, -karthik From skadron at cs.virginia.edu Tue May 8 21:11:04 2007 From: skadron at cs.virginia.edu (Kevin Skadron) Date: Tue May 8 21:10:29 2007 Subject: [Hotspot] Re: Hotspot Query In-Reply-To: References: <457828DC.6090006@mail.cs.virginia.edu> <592D759C9E9D96196E8D2117@mstu2> <96F600E1A1994187CE305A89@mstu2> <511C87CF989B30C59017CE8B@mstu2> <3967F151E6930F56B8CABB78@mstu2> <8707E5418AEC52D23FD00E96@mstu2> <3B4C9497E4863D4C01D2BC31@mstu2> <9B6257EE5E9AFE20B844A4A1@mstu2> Message-ID: <46411FA8.2070607@mail.cs.virginia.edu> This raises an interesting point that we never really paid close attention (or at least, I didn't) to the UIUC WDDD paper a couple years ago about HotSpot. Is there anything in there that we need to deal with? If Sarita is at WDDD this year, she might even ask about this, so Karthik, you should look into this. /K Wei Huang wrote: > Hi Niti, glad we can help! Please see my replies below. (Karthik may > have more insights.) > > -Wei > >> 1) Typically, what is the recommended temperature metric >> to report in papers - steady-state temperatures or >> the average of transient temperature? I have seen >> some papers where the authors advocate running Hotspot >> twice after initiliazing the second run with first run's >> steady-state temperatures. But they report the second >> run's steady-state temperature. Shouldn't the >> steady-state temperatures be the same in both runs? I >> see only transient temperatures to be different in 2 runs. > > I can't see why 2 runs are needed for steady temperatures. Can you point > me to the papers you mentioned? Normally, steady-state temperatures are > useful for initializing transient thermal simulations. But you still > need some additional "warm-up" simulations after setting initial > temperatures to steady-state ones. Steady-state temperature is also > useful when transient power numbers are not available or too expensive > to obtain. I guess this is not your case, but for circuit/SOC/ASIC > designers, this is usually true. Because for them, doing > clock-cycle-accurate circuit simulations to model transient temperature > needs a super powerful and detailed transient power estimator and super > long simulation time. > >> >> 2) For a 5 us sampling interval, the transient >> temperature for hotter blocks is higher than the >> steady-state temperature in the second run. I saw a few >> mails in your online archived mailing list about >> steady-state temperature not reaching. According to >> that, a much larger sampling interval is recommended for >> transient and steady-state results to match. Is there any >> recommended value for this interval? > > In HotSpot, we use average power to calculate steady temperatures. So it > is normal to see transient temperatures higher or lower than steady > ones. Using a large sampling interval is essentially equivalent to > averaging the power trace over a longer time, so it is not surprising to > see that this matches the steady temperature better. By and large, you > can trust your transient simulations (provided with a long enough > "warm-up" stage). Steady temperature is more of a "reference" that can > be useful in some cases as I mentioned before. > >> >> 3) I wanted to model DTM in my implementation. I need to >> use transient temperatures to trigger the DTM in the >> second run. I saw a paper in WDDD: >> >> http://www.ece.wisc.edu/~wddd/2005/papers/WDDD05_srinivas >> an.pdf According to this paper, their tool as well as >> Hotspot doesn't model reactive DTM if used with only 2 >> runs as the heat-sink model is not initialized >> adaptively for DTM. They recommend 5-6 runs to model >> reactive DTM. Do you agree with this methodology? > > Again, I think this is related to the "warm-up". The warm-up stage has > to be long enough to "heat up" the heatsink. So you may want to monitor > the heat sink temperature to see if it stabilizes (i.e. some of the > "EXTRA" nodes in HotSpot code). However, you may also try to tweak the > thermal capacitance of the heatsink to small values during the warm-up > phase, so that it artificially brings the heat sink to a steady > temperature much faster. But you need to be really careful with this > since it is error-prone. I don't recommend this unless you are really > familiar with the HotSpot code base and how it works. >> >> 4) Finally, the DTM macro in Hotspot, does it model >> heat-sink initialization adaptively if enabled? I have >> never used it and wasn't sure how it works. > > I am not 100% percent sure. Karthik should be more helpful than I am on > this issue. But I assume this is the case since the model itself already > has all the structures needed for this. Hope this helps. > >> >> Thanks for all your help. >> >> -Niti >> > > > > > > _______________________________________________ > HotSpot mailing list > HotSpot@ares.cs.Virginia.EDU > http://www.cs.Virginia.EDU/mailman-2.1.5/listinfo/hotspot From skadron at cs.virginia.edu Tue May 8 21:14:59 2007 From: skadron at cs.virginia.edu (Kevin Skadron) Date: Tue May 8 21:14:23 2007 Subject: [Hotspot] Re: Hotspot Query In-Reply-To: <46411FA8.2070607@mail.cs.virginia.edu> References: <457828DC.6090006@mail.cs.virginia.edu> <96F600E1A1994187CE305A89@mstu2> <511C87CF989B30C59017CE8B@mstu2> <3967F151E6930F56B8CABB78@mstu2> <8707E5418AEC52D23FD00E96@mstu2> <3B4C9497E4863D4C01D2BC31@mstu2> <9B6257EE5E9AFE20B844A4A1@mstu2> <46411FA8.2070607@mail.cs.virginia.edu> Message-ID: <46412093.2070109@mail.cs.virginia.edu> Whoops Niti, I was sloppy with my CC's! (In case you're curious, we *did* discuss this but felt that we had already dealt with this in our earlier paper, but in the midst of other things, never popped the stack to look more closely. Now that we have re-examined the heat sink treatment in our upcoming WDDD paper, we should probably go back and look at their paper more closely.) /K Kevin Skadron wrote: > This raises an interesting point that we never really paid close > attention (or at least, I didn't) to the UIUC WDDD paper a couple years > ago about HotSpot. Is there anything in there that we need to deal with? > > If Sarita is at WDDD this year, she might even ask about this, so > Karthik, you should look into this. > > /K > > Wei Huang wrote: >> Hi Niti, glad we can help! Please see my replies below. (Karthik may >> have more insights.) >> >> -Wei >> >>> 1) Typically, what is the recommended temperature metric >>> to report in papers - steady-state temperatures or >>> the average of transient temperature? I have seen >>> some papers where the authors advocate running Hotspot >>> twice after initiliazing the second run with first run's >>> steady-state temperatures. But they report the second >>> run's steady-state temperature. Shouldn't the >>> steady-state temperatures be the same in both runs? I >>> see only transient temperatures to be different in 2 runs. >> >> I can't see why 2 runs are needed for steady temperatures. Can you >> point me to the papers you mentioned? Normally, steady-state >> temperatures are useful for initializing transient thermal >> simulations. But you still need some additional "warm-up" simulations >> after setting initial temperatures to steady-state ones. Steady-state >> temperature is also useful when transient power numbers are not >> available or too expensive to obtain. I guess this is not your case, >> but for circuit/SOC/ASIC designers, this is usually true. Because for >> them, doing clock-cycle-accurate circuit simulations to model >> transient temperature needs a super powerful and detailed transient >> power estimator and super long simulation time. >> >>> >>> 2) For a 5 us sampling interval, the transient >>> temperature for hotter blocks is higher than the >>> steady-state temperature in the second run. I saw a few >>> mails in your online archived mailing list about >>> steady-state temperature not reaching. According to >>> that, a much larger sampling interval is recommended for >>> transient and steady-state results to match. Is there any >>> recommended value for this interval? >> >> In HotSpot, we use average power to calculate steady temperatures. So >> it is normal to see transient temperatures higher or lower than steady >> ones. Using a large sampling interval is essentially equivalent to >> averaging the power trace over a longer time, so it is not surprising >> to see that this matches the steady temperature better. By and large, >> you can trust your transient simulations (provided with a long enough >> "warm-up" stage). Steady temperature is more of a "reference" that can >> be useful in some cases as I mentioned before. >> >>> >>> 3) I wanted to model DTM in my implementation. I need to >>> use transient temperatures to trigger the DTM in the >>> second run. I saw a paper in WDDD: >>> >>> http://www.ece.wisc.edu/~wddd/2005/papers/WDDD05_srinivas >>> an.pdf According to this paper, their tool as well as >>> Hotspot doesn't model reactive DTM if used with only 2 >>> runs as the heat-sink model is not initialized >>> adaptively for DTM. They recommend 5-6 runs to model >>> reactive DTM. Do you agree with this methodology? >> >> Again, I think this is related to the "warm-up". The warm-up stage has >> to be long enough to "heat up" the heatsink. So you may want to >> monitor the heat sink temperature to see if it stabilizes (i.e. some >> of the "EXTRA" nodes in HotSpot code). However, you may also try to >> tweak the thermal capacitance of the heatsink to small values during >> the warm-up phase, so that it artificially brings the heat sink to a >> steady temperature much faster. But you need to be really careful with >> this since it is error-prone. I don't recommend this unless you are >> really familiar with the HotSpot code base and how it works. >>> >>> 4) Finally, the DTM macro in Hotspot, does it model >>> heat-sink initialization adaptively if enabled? I have >>> never used it and wasn't sure how it works. >> >> I am not 100% percent sure. Karthik should be more helpful than I am >> on this issue. But I assume this is the case since the model itself >> already has all the structures needed for this. Hope this helps. >> >>> >>> Thanks for all your help. >>> >>> -Niti >>> >> >> >> >> >> >> _______________________________________________ >> HotSpot mailing list >> HotSpot@ares.cs.Virginia.EDU >> http://www.cs.Virginia.EDU/mailman-2.1.5/listinfo/hotspot > > From skadron at cs.virginia.edu Fri May 11 09:36:45 2007 From: skadron at cs.virginia.edu (Kevin Skadron) Date: Fri May 11 09:36:34 2007 Subject: [Hotspot] HOTSPOT with benchmarks In-Reply-To: <507e82580705071535q76a95731q41be355738b91b22@mail.gmail.com> References: <507e82580705071535q76a95731q41be355738b91b22@mail.gmail.com> Message-ID: <4644716D.8020302@mail.cs.virginia.edu> Nishchay Bharati wrote: > Hi Everyone, > I have just started using HOTSPOT and I wanted to > know how I can run HOTSPOT with SPEC2006 benchmarks. I mean how should I > modify the command line for HOTSPOT to see how the different units of > the processor heat up to different benchmarks. Nishchay, What you need to do is identify the command line that each of those benchmarks needs. You'll need this for any use of those benchmarks, not just HotSpot (unless you only run the benchmarks within the SPEC harness, and in that case, you would be integrating the simulator into those SPEC tools as an additional step). You should be able to extract the command lines by examining how the SPEC tools invoke each benchmark. A compilers group here at U.Va. has gone through this process for more of the benchmarks than my group, and they kindly provided the following info. I think this covers all the integer benchmarks and most but not all of the FP. /K Integer, Test Inputs astar lake.cfg bzip2 input.program 5 bzip2 dryer.jpg 2 gcc cccp.i -o cccp.s gobmk --quiet --mode gtp < capture.tst gobmk --quiet --mode gtp < connect.tst gobmk --quiet --mode gtp < connect_rot.tst gobmk --quiet --mode gtp < connection.tst gobmk --quiet --mode gtp < connection_rot.tst gobmk --quiet --mode gtp < cutstone.tst gobmk --quiet --mode gtp < dniwog.tst h264ref -d foreman_test_encoder_baseline.cfg hmmer --fixed 0 --mean 325 --num 45000 --sd 200 --seed 0 bombesin.hmm libquantum 33 5 mcf inp.in omnetpp omnetpp.ini perlbench -I. -Ilib attrs.pl perlbench -I. -Ilib gv.pl perlbench -I. -Ilib makerand.pl perlbench -I. -Ilib pack.pl perlbench -I. -Ilib redef.pl perlbench -I. -Ilib ref.pl perlbench -I. -Ilib regmsg.pl perlbench -I. -Ilib test.pl sjeng test.txt Xalan -v test.xml xalanc.xsl Integer, Reference Inputs astar BigLakes2048.cfg astar rivers.cfg bzip2 input.source 280 bzip2 chicken.jpg 30 bzip2 liberty.jpg 30 bzip2 input.program 280 bzip2 text.html 280 bzip2 input.combined 200 gcc 166.i -o 166.s gcc 200.i -o 200.s gcc c-typeck.i -o c-typeck.s gcc cp-decl.i -o cp-decl.s gcc expr.i -o expr.s gcc expr2.i -o expr2.s gcc g23.i -o g23.s gcc s04.i -o s04.s gcc scilab.i -o scilab.s gobmk --quiet --mode gtp < 13x13.tst gobmk --quiet --mode gtp < nngs.tst gobmk --quiet --mode gtp < score2.tst gobmk --quiet --mode gtp < trevorc.tst gobmk --quiet --mode gtp < trevord.tst h264ref -d foreman_ref_encoder_baseline.cfg h264ref -d foreman_ref_encoder_main.cfg h264ref -d sss_encoder_main.cfg libquantum 1397 8 mcf inp.in omnetpp omnetpp.ini perlbench -Ilib checkspam.pl 2500 5 25 11 150 1 1 1 1 perlbench -Ilib diffmail.pl 4 800 10 17 19 300 perlbench -Ilib splitmail.pl 1600 12 26 16 4500 sjeng ref.txt Xalan -v t5.xml xalanc.xsl Floating Point, Test Inputs dealII 8 GemsFDTD lbm 20 reference.dat 0 1 100_100_130_cf_a.of leslie3d < leslie3d.in milc < su3imp.in namd --input namd.input --iterations 1 --output namd.out povray SPEC-benchmark-test.ini soplex -m10000 test.mps sphinx_livepretend ctlfile . args.an4 tonto zeusmp Floating Point, Reference Inputs dealII 23 GemsFDTD lbm 3000 reference.dat 0 0 100_100_130_ldc.of leslie3d < leslie3d.in milc < su3imp.in namd --input namd.input --iterations 38 --output namd.out povray SPEC-benchmark-ref.ini soplex -s1 -e -m45000 pds-50.mps soplex -m3500 ref.mps sphinx_livepretend ctlfile . args.an4 tonto zeusmp