From bernauer at informatik.uni-tuebingen.de  Mon Mar 10 01:05:37 2008
From: bernauer at informatik.uni-tuebingen.de (Andreas Bernauer)
Date: Mon, 10 Mar 2008 10:05:37 +0100
Subject: [Hotspot] Reduce simulation time
Message-ID: <1205139937.10930.63.camel@lava.informatik.uni-tuebingen.de>

Hi,

first, thank you for your great software.  It helps us a lot in our
research.

We integrated Hotspot 4.0 in our simulator (of a System-on-Chip, SoC).
Although I've enabled Hotspot's math acceleration for my CPU, the
simulator still spends 90% of its time within Hotspot estimating
temperatures.  I hope you can help me to improve simulation speed.

Turning on verbosity tells me that rk4 is called ~5200 times during
compute_temp.  Is this typical for a simulation time of 1s?

I tried to reduce computation time by increasing MIN_STEP in
temperature.h and decreasing RK4_PRECISION in RCutil.c, to no avail.  I
printed out the intermediate temperatures in the for-loop in
compute_temp_block in temperature_block.c and found that the
temperatures are oscillating (see attached plot).  Maybe this is the
reason for the long computation time?

Do you know a way of reducing computation time, even at the price of
reduced precision?

Thank you very much in advance for your answer.

Best regards,

Andreas.

-- 
WSI/TI, Sand 13, 72076 T?bingen, +49 70 71 29 75 940
http://www.ti.uni-tuebingen.de/Andreas_Bernauer.227.0.html

-------------- next part --------------
A non-text attachment was scrubbed...
Name: temp.png
Type: application/octet-stream
Size: 6328 bytes
Desc: not available
Url : http://www.cs.virginia.edu/pipermail/hotspot/attachments/20080310/cbe5d8ee/attachment.obj 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : http://www.cs.virginia.edu/pipermail/hotspot/attachments/20080310/cbe5d8ee/attachment.bin 

From wh6p at cms.mail.virginia.edu  Mon Mar 10 06:52:06 2008
From: wh6p at cms.mail.virginia.edu (Wei Huang)
Date: Mon, 10 Mar 2008 10:52:06 -0400
Subject: [Hotspot] Reduce simulation time
In-Reply-To: <1205139937.10930.63.camel@lava.informatik.uni-tuebingen.de>
References: <1205139937.10930.63.camel@lava.informatik.uni-t
 uebingen.de>
Message-ID: <8E738840525FE7D8D160D47E@mstu2>

Hi Andreas,

Are you running HotSpot on a per-clock-cycle basis or using 
the default sampling interval in hotspot.config? You may 
want to sample temperatures less frequently and see if that 
affects the accuracy a lot.

Other factors that could have impact on simulation speed 
are the number of blocks you have in the floorplan and 
number of blocks with high aspect ratios.

Thanks,
-Wei

--On Monday, March 10, 2008 10:05 AM +0100 Andreas Bernauer 
<bernauer at informatik.uni-tuebingen.de> wrote:

> Hi,
>
> first, thank you for your great software.  It helps us a
> lot in our research.
>
> We integrated Hotspot 4.0 in our simulator (of a
> System-on-Chip, SoC). Although I've enabled Hotspot's
> math acceleration for my CPU, the simulator still spends
> 90% of its time within Hotspot estimating temperatures.
> I hope you can help me to improve simulation speed.
>
> Turning on verbosity tells me that rk4 is called ~5200
> times during compute_temp.  Is this typical for a
> simulation time of 1s?
>
> I tried to reduce computation time by increasing MIN_STEP
> in temperature.h and decreasing RK4_PRECISION in
> RCutil.c, to no avail.  I printed out the intermediate
> temperatures in the for-loop in compute_temp_block in
> temperature_block.c and found that the temperatures are
> oscillating (see attached plot).  Maybe this is the
> reason for the long computation time?
>
> Do you know a way of reducing computation time, even at
> the price of reduced precision?
>
> Thank you very much in advance for your answer.
>
> Best regards,
>
> Andreas.
>
> --
> WSI/TI, Sand 13, 72076 T?bingen, +49 70 71 29 75 940
> http://www.ti.uni-tuebingen.de/Andreas_Bernauer.227.0.html
>







From ks4kk at cs.virginia.edu  Mon Mar 10 09:45:30 2008
From: ks4kk at cs.virginia.edu (Karthik Sankaranarayanan)
Date: Mon, 10 Mar 2008 13:45:30 -0400 (EDT)
Subject: [Hotspot] Reduce simulation time
In-Reply-To: <1205139937.10930.63.camel@lava.informatik.uni-tuebingen.de>
References: <1205139937.10930.63.camel@lava.informatik.uni-tuebingen.de>
Message-ID: <Pine.SOC.4.64.0803101318350.6230@cobra.cs.Virginia.EDU>


Hello Andreas,

> Turning on verbosity tells me that rk4 is called ~5200 times during
> compute_temp.  Is this typical for a simulation time of 1s?

Yes - it looks OK to me. With adaptive RK4, this averages to about 0.2 ms 
per RK4 step, which is about right for a 1s interval. If you want RK4 to 
adapt faster (since your calling interval is large - 1s), you might want 
to relax the RK4_MAXUP (in RCutil.c) a litte bit. This increases the upper 
limit on how much the step size can increase in each step. So, the 
adaptation can happen quicker.

> I tried to reduce computation time by increasing MIN_STEP in
> temperature.h and decreasing RK4_PRECISION in RCutil.c, to no avail.  I
> printed out the intermediate temperatures in the for-loop in
> compute_temp_block in temperature_block.c and found that the
> temperatures are oscillating (see attached plot).  Maybe this is the
> reason for the long computation time?

I am not sure if I understand the plot fully but here is my take. It seems 
that temperatures of three different units have been plotted. The small 
oscillatory features (jagged lines) seem normal to me because they are 
well below the desired RK4_PRECISION, which is 0.01 degrees. Zooming out 
the plot to that level (> 5 times) would mean pretty smooth curves.

> Do you know a way of reducing computation time, even at the price of
> reduced precision?

Apart from the RK4_MAXUP suggestion, a couple more come to mind:

a) How many blocks are there in the floorplan? Depending on this number, 
you might speed-up the simulation by switching to the grid model with a 
coarser grid.

b) While using the grid model, you could also simplify the modeling of the 
package by removing the TIM layer through the layer configuration file 
specification.

Hope this helps.
-karthik

From bernauer at informatik.uni-tuebingen.de  Tue Mar 11 08:30:06 2008
From: bernauer at informatik.uni-tuebingen.de (bernauer at informatik.uni-tuebingen.de)
Date: Tue, 11 Mar 2008 17:30:06 +0100 (MET)
Subject: [Hotspot] Reduce simulation time
In-Reply-To: <Pine.SOC.4.64.0803101318350.6230@cobra.cs.Virginia.EDU>
References: <1205139937.10930.63.camel@lava.informatik.uni-tuebingen.de>
	<Pine.SOC.4.64.0803101318350.6230@cobra.cs.Virginia.EDU>
Message-ID: <42349.134.2.12.41.1205253006.squirrel@imap>

Hello Karthik, Hello Wei,

thank you very much for your responses.  Sorry that I missed some
explanations. The plots are temperature curves of three different units. 
I have 6 blocks, one of which has a rather high aspect ratio, so I could
split this one into two blocks. I am running the simulation at the same
time I am gaining the power estimates, collecting the power over 1s, then
running Hotspot.  Currently, I don't have a way to change the sampling
frequency of 1s.

Thank you Karthik for telling me that my plots are "normal" :-) I'll try
to relax RK4_MAXUP and to use the grid model without TIM layer.

BTW: If the thermal resistance between all units is the same, could the
calculation be simplified (ie. replace rk4 by something simpler/faster)? 
Or did I get the model wrong?

Best regards,

Andreas.




From gokulv at mail.utexas.edu  Thu Mar 13 11:26:59 2008
From: gokulv at mail.utexas.edu (gokulv at mail.utexas.edu)
Date: Thu, 13 Mar 2008 14:26:59 -0500
Subject: [Hotspot] query
Message-ID: <1205436419.47d98003f3615@webmailapp6.cc.utexas.edu>

Hi,

I am a user of Hotspot. I was doing some preliminary simulations.
I changed the -sampling_intvl to some other value and reran the simulation with
the -d option. The value specified in the config file and output of -d option
file do not match.

I am using the version 4 of the tool. Am I missing something.

options given:
# hotspot calling interval
	-sampling_intvl		0.0105e-04


Thanks for your help in advance.

Regards
Gokul


From wh6p at cms.mail.virginia.edu  Fri Mar 14 04:51:48 2008
From: wh6p at cms.mail.virginia.edu (Wei Huang)
Date: Fri, 14 Mar 2008 08:51:48 -0400
Subject: [Hotspot] query
In-Reply-To: <1205436419.47d98003f3615@webmailapp6.cc.utexas.edu>
References: <1205436419.47d98003f3615@webmailapp6.cc.utexas.
 edu>
Message-ID: <CBCDBB1C58C933F8415D328A@[192.168.1.102]>

Hi Gokul,

Did you also use the -c option in the command line?

-Wei

--On Thursday, March 13, 2008 2:26 PM -0500 
gokulv at mail.utexas.edu wrote:

> Hi,
>
> I am a user of Hotspot. I was doing some preliminary
> simulations. I changed the -sampling_intvl to some other
> value and reran the simulation with the -d option. The
> value specified in the config file and output of -d option
> file do not match.
>
> I am using the version 4 of the tool. Am I missing
> something.
>
> options given:
># hotspot calling interval
> 	-sampling_intvl		0.0105e-04
>
>
> Thanks for your help in advance.
>
> Regards
> Gokul
>
> _______________________________________________
> HotSpot mailing list
> HotSpot at mail.cs.virginia.edu
> http://www.cs.virginia.edu/mailman/listinfo/hotspot






From michaelkadin at gmail.com  Fri Mar 14 05:19:30 2008
From: michaelkadin at gmail.com (Mike Kadin)
Date: Fri, 14 Mar 2008 09:19:30 -0400
Subject: [Hotspot] Difference between Steady State and Transient
Message-ID: <34954c9d0803140619u5476bb44gf6701e8ed6605ede@mail.gmail.com>

Hey All,

I've been having some trouble with hotspot that I haven't previously
encountered and was hoping for a little guidance, advice.


I'm using a pretty standard floorplan, and power traces I've pulled from
PTScalar.  I'm using the suggested method of running HotSpot 2 times, once
for steady state values, and once for transient values.  The steady state
block temperatures seem to be in a reasonable range (75-150 degrees C).
When plotting the transient data for each block, I can see it does level off
over time, but for a few blocks (not all of them) the graph levels off at
values closer to 350 degrees C.

Any ideas on why the steady-state and transient data could differ so much?

Any help would be greatly appreciated.

Thanks,

Mike Kadin

Undergraduate Researcher
Brown University Division of Engineering
kadin at brown.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.cs.virginia.edu/pipermail/hotspot/attachments/20080314/e9b489f0/attachment.html 

From wh6p at cms.mail.virginia.edu  Fri Mar 14 06:42:17 2008
From: wh6p at cms.mail.virginia.edu (Wei Huang)
Date: Fri, 14 Mar 2008 10:42:17 -0400
Subject: [Hotspot] Difference between Steady State and Transient
In-Reply-To: <34954c9d0803140619u5476bb44gf6701e8ed6605ede@mail.gmail.com>
References: <34954c9d0803140619u5476bb44gf6701e8ed6605ede@ma
 il.gmail.com>
Message-ID: <A5EC638FEABBE212A06F1867@mstu2>

Hi Mike,

Do you have any overlapping units in the floorplan or any 
uncommon blocks such as ones with very high aspect ratio or 
very high power density etc? Another HotSpot user 
experienced similar problems last year, and it turned out 
that the floorplan had overlapping units. Please also tell 
us which version of HotSpot are you using? thanks.

-Wei

--On Friday, March 14, 2008 9:19 AM -0400 Mike Kadin 
<michaelkadin at gmail.com> wrote:

> Hey All,
>
> I've been having some trouble with hotspot that I haven't
> previously encountered and was hoping for a little
> guidance, advice.
>
>
> I'm using a pretty standard floorplan, and power traces
> I've pulled from PTScalar.  I'm using the suggested
> method of running HotSpot 2 times, once for steady state
> values, and once for transient values.  The steady state
> block temperatures seem to be in a reasonable range
> (75-150 degrees C).  When plotting the transient data for
> each block, I can see it does level off over time, but
> for a few blocks (not all of them) the graph levels off
> at values closer to 350 degrees C.
>
> Any ideas on why the steady-state and transient data
> could differ so much?
>
> Any help would be greatly appreciated.
>
> Thanks,
>
> Mike Kadin
>
> Undergraduate Researcher
> Brown University Division of Engineering
> kadin at brown.edu
>
>






From gokulv at mail.utexas.edu  Fri Mar 14 14:10:03 2008
From: gokulv at mail.utexas.edu (gokulv at mail.utexas.edu)
Date: Fri, 14 Mar 2008 17:10:03 -0500
Subject: [Hotspot] query
In-Reply-To: <CBCDBB1C58C933F8415D328A@[192.168.1.102]>
References: <1205436419.47d98003f3615@webmailapp6.cc.utexas. edu>
	<CBCDBB1C58C933F8415D328A@[192.168.1.102]>
Message-ID: <1205532603.47daf7bb113a1@webmailapp6.cc.utexas.edu>

Hi Wei,

Thanks for replying. I was able the fix the issue with the help of Karthik.

Thanks and Regards,
Gokul


Quoting Wei Huang <wh6p at cms.mail.virginia.edu>:

> Hi Gokul,
>
> Did you also use the -c option in the command line?
>
> -Wei
>
> --On Thursday, March 13, 2008 2:26 PM -0500
> gokulv at mail.utexas.edu wrote:
>
> > Hi,
> >
> > I am a user of Hotspot. I was doing some preliminary
> > simulations. I changed the -sampling_intvl to some other
> > value and reran the simulation with the -d option. The
> > value specified in the config file and output of -d option
> > file do not match.
> >
> > I am using the version 4 of the tool. Am I missing
> > something.
> >
> > options given:
> ># hotspot calling interval
> > 	-sampling_intvl		0.0105e-04
> >
> >
> > Thanks for your help in advance.
> >
> > Regards
> > Gokul
> >
> > _______________________________________________
> > HotSpot mailing list
> > HotSpot at mail.cs.virginia.edu
> > http://www.cs.virginia.edu/mailman/listinfo/hotspot
>
>
>
>
>




From wh6p at cms.mail.virginia.edu  Thu Mar 27 12:26:19 2008
From: wh6p at cms.mail.virginia.edu (Wei Huang)
Date: Thu, 27 Mar 2008 16:26:19 -0400
Subject: [Hotspot] A problem related to HotSpot 4.0 (fwd)
Message-ID: <3A6323CDC40896FECF61F3A8@mstu2>

Wei Huang wrote:
> Just to clarify, HotSpot grid model is able to model
> multiple layers  with power dissipations like 3D chips.
> In order to do this, you need to  use the grid model by
> setting "-model_type grid", and modify the "lcf"  layer
> configuration file--just follow the instructions in the
> sample lcf  file. Let us know if it works. Thanks.
>
> -Wei
>
>> On Thu, 27 Mar 2008 20:04:06 +0530
>>   "Sudarshan Reddy M" <sudarshanml at gmail.com> wrote:
>>> Hi,
>>> I am using HotSpot 4.0 for thermal estimation of 3D
>>> Chips. It seems there is a problem in reading power
>>> trace file and populating it into the HotSpot internal
>>> vectors. I have used 2 layer stack with an insulation
>>> layer in between which doesn't dissipate power. What I
>>> have noticed is that the power in the second active
>>> layer is not populated in the "overall_power", thus
>>> resulting in incorrect temperatures. Please let me know
>>> if I am missing anything or it is indeed a problem.
>>> Regards
>>> Sudarshan Reddy M
>>> IIT Bombay



From ks4kk at cs.virginia.edu  Fri Mar 28 22:37:00 2008
From: ks4kk at cs.virginia.edu (Karthik Sankaranarayanan)
Date: Sat, 29 Mar 2008 02:37:00 -0400 (EDT)
Subject: [Hotspot] A problem related to HotSpot 4.0
In-Reply-To: <fa7c83050803280818s57fbdadeub191b6692acd3675@mail.gmail.com>
References: <web-2613641@guppy.mail.virginia.edu>
	<873B508C66CF31960460DCCA@192.168.1.102>
	<fa7c83050803280247j3b3db697tcfd779b9380cfe8b@mail.gmail.com> 
	<F5172994A5BB76907EF4D28C@mstu2>
	<fa7c83050803280818s57fbdadeub191b6692acd3675@mail.gmail.com>
Message-ID: <Pine.SOC.4.64.0803290141370.25475@cobra.cs.Virginia.EDU>


Whoops - this is indeed a bug in HotSpot! Thank you very much for letting 
us know Sudarshan. Here is a patch file that fixes this issue related to 
3D chip modeling. Please apply it to the hotspot.c file of the 4.0 
distribution (save the patchfile in the HotSpot-4.0 directory; cd 
HotSpot-4.0; patch hotspot.c patchfile). The next release of HotSpot will 
include this fix.

Hope this answers your question. Thank you again.

Sincerely,
-karthik


On Fri, 28 Mar 2008, Sudarshan Reddy M wrote:

> Hi,
> I have specified the lcf file in .config and HotSpot is indeed dumping
> all the layer temperatures.
> I debugged the problem and found that when "overall_power" (can be
> found in hotspot.c) populated it is not done correctly. What I have
> observed is that vector called "power" is being populated from the
> power trace file  and then "overall_power" is calculated as the
> average power. Here "power" will have non-zero power 0 - n1 where n1
> is the total number of units in the first layer and (n1+1)-(n2) will
> be all zeros (because I have specified second layer as a
> non-dissipative one) and (n2+1)-n3 will represent the third power
> where n3 is the total number of units in the third layer. However
> while populating "overall_power" it is done only from 0 to n4 where n4
> is the number of units in all the power dissipating layers, which
> seems to be wrong(might be a bug in the software). What I am not sure
> is whether that is indeed a problem or not?
> It will be more than helpful if you can comment on this.
> Thank you very much for the quick responses and the time you have
> spent in answering my queries.
>
> Regards
> Sudarshan.
-------------- next part --------------
--- ../HotSpot-4.0/hotspot.c	2007-09-02 01:23:02.000000000 -0400
+++ hotspot.c	2008-03-29 01:31:41.695717000 -0400
@@ -449,8 +449,17 @@
 		}		
 	
 		/* for computing average	*/
-		for(i=0; i < n; i++)
-			overall_power[i] += power[i];
+		if (model->type == BLOCK_MODEL)
+			for(i=0; i < n; i++)
+				overall_power[i] += power[i];
+		else
+			for(i=0, base=0; i < model->grid->n_layers; i++) {
+				if(model->grid->layers[i].has_power)
+					for(j=0; j < model->grid->layers[i].flp->n_units; j++)
+						overall_power[base+j] += power[base+j];
+				base += model->grid->layers[i].flp->n_units;	
+			}
+
 		lines++;
 	}
 
@@ -458,8 +467,16 @@
 		fatal("no power numbers in trace file\n");
 		
 	/* for computing average	*/
-	for(i=0; i < n; i++)
-		overall_power[i] /= lines;
+	if (model->type == BLOCK_MODEL)
+		for(i=0; i < n; i++)
+			overall_power[i] /= lines;
+	else
+		for(i=0, base=0; i < model->grid->n_layers; i++) {
+			if(model->grid->layers[i].has_power)
+				for(j=0; j < model->grid->layers[i].flp->n_units; j++)
+					overall_power[base+j] /= lines;
+			base += model->grid->layers[i].flp->n_units;	
+		}
 
 	/* steady state temperature	*/
 	steady_state_temp(model, overall_power, steady_temp);

From sudarshanml at gmail.com  Fri Mar 28 23:32:16 2008
From: sudarshanml at gmail.com (Sudarshan Reddy M)
Date: Sat, 29 Mar 2008 13:02:16 +0530
Subject: [Hotspot] A problem related to HotSpot 4.0
In-Reply-To: <Pine.SOC.4.64.0803290141370.25475@cobra.cs.Virginia.EDU>
References: <web-2613641@guppy.mail.virginia.edu>
	<873B508C66CF31960460DCCA@192.168.1.102>
	<fa7c83050803280247j3b3db697tcfd779b9380cfe8b@mail.gmail.com>
	<F5172994A5BB76907EF4D28C@mstu2>
	<fa7c83050803280818s57fbdadeub191b6692acd3675@mail.gmail.com>
	<Pine.SOC.4.64.0803290141370.25475@cobra.cs.Virginia.EDU>
Message-ID: <fa7c83050803290032r3bfde69fy537e5cc5fbcff962@mail.gmail.com>

Hi Karthik,
Thank you very much for the clarification and the fix. My sincere
thanks to all the people in your group not only for these quick
responses but also for the beautiful tool you people have provided to
the user community.
Thanks
Sudarshan.

On Sat, Mar 29, 2008 at 12:07 PM, Karthik Sankaranarayanan
<ks4kk at cs.virginia.edu> wrote:
>
>  Whoops - this is indeed a bug in HotSpot! Thank you very much for letting
>  us know Sudarshan. Here is a patch file that fixes this issue related to
>  3D chip modeling. Please apply it to the hotspot.c file of the 4.0
>  distribution (save the patchfile in the HotSpot-4.0 directory; cd
>  HotSpot-4.0; patch hotspot.c patchfile). The next release of HotSpot will
>  include this fix.
>
>  Hope this answers your question. Thank you again.
>
>  Sincerely,
>  -karthik
>
>
>
>
>  On Fri, 28 Mar 2008, Sudarshan Reddy M wrote:
>
>  > Hi,
>  > I have specified the lcf file in .config and HotSpot is indeed dumping
>  > all the layer temperatures.
>  > I debugged the problem and found that when "overall_power" (can be
>  > found in hotspot.c) populated it is not done correctly. What I have
>  > observed is that vector called "power" is being populated from the
>  > power trace file  and then "overall_power" is calculated as the
>  > average power. Here "power" will have non-zero power 0 - n1 where n1
>  > is the total number of units in the first layer and (n1+1)-(n2) will
>  > be all zeros (because I have specified second layer as a
>  > non-dissipative one) and (n2+1)-n3 will represent the third power
>  > where n3 is the total number of units in the third layer. However
>  > while populating "overall_power" it is done only from 0 to n4 where n4
>  > is the number of units in all the power dissipating layers, which
>  > seems to be wrong(might be a bug in the software). What I am not sure
>  > is whether that is indeed a problem or not?
>  > It will be more than helpful if you can comment on this.
>  > Thank you very much for the quick responses and the time you have
>  > spent in answering my queries.
>  >
>  > Regards
>  > Sudarshan.



-- 
Sudarshan Reddy M
M. Tech Control and Computing
IIT Bombay