Re: [Gems-users] Understand the protocol trace

["Lei Yang" <lya755@xxxxxxxxxxxxxxxxxxxx>]

I changed the network topology to PT_TO_PT and the cycle counts reduced. 
However I found a wierd problem: when I set g_NUM_PROCESSORS to 16 and 
g_PROCS_PER_CHIP to 1, 2, 4, or 8, the tester work fine; but when I set it 
to other values it reports assertion failure.
For example, I tried the following combinations:

x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 1 -a 1 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 4 -a 1 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 4 -a 2 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 4 -a 4 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 8 -a 1 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 8 -a 2 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 8 -a 4 -z 
little.trace
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 8 -a 8 -z 
little.trace
They gave me this error:

Testing clear stats...Done.
Reading trace from file 'little.trace'...
failed assertion 'index < m_size' at fn const TYPE& Vector<TYPE>::ref(int) 
const [with TYPE = AbstractChip*] in ../common/Vector.h:157
failed assertion 'index < m_size' at fn const TYPE& Vector<TYPE>::ref(int) 
const [with TYPE = AbstractChip*] in ../common/Vector.h:157
At this point you might want to attach a debug to the running and get to the
crash site; otherwise press enter to continue
PID: 3889

I also tried this:
x86-linux/generated/MSI_MOSI_CMP_directory/bin/tester.exec -p 16 -a 16 -z 
little.trace
This gave me a seg fault:

failed assertion 'L2_CACHE_NUM_SETS_BITS >= 
log_int(g_NUM_L2_BANKS_PER_CHIP)' at fn static void RubyConfig::init() in 
config/RubyConfig.C:123
Segmentation fault

Could somebody give me a hint on what is the problem? Thanks a lot!!

Lei
----- Original Message ----- 
From: "Lei Yang" <lya755@xxxxxxxxxxxxxxxxxxxx>
To: "Mike Marty" <mikem@xxxxxxxxxxx>; "Gems Users" <gems-users@xxxxxxxxxxx>
Sent: Monday, September 25, 2006 12:06 PM
Subject: Re: [Gems-users] Understand the protocol trace


> I just found in the documentation that PT_TO_PT and FILE_SPECIFIED are the
> recommended network topologies for the CMP protocols.  Will give that a 
> try.
> Lei
> ----- Original Message ----- 
> From: "Mike Marty" <mikem@xxxxxxxxxxx>
> To: "Lei Yang" <lya755@xxxxxxxxxxxxxxxxxxxx>; "Gems Users"
> <gems-users@xxxxxxxxxxx>
> Sent: Monday, September 25, 2006 11:59 AM
> Subject: Re: [Gems-users] Understand the protocol trace
>
>
> > 1)  Turn off RANDOMIZATION in $GEMS/ruby/config/testerconfig.defaults.
> > This randomly adds 100+ cycle delays to generate race conditions
> >
> > 2)  You are probably using a non-CMP topology and NETWORK_LINK_LATENCY is
> > fairly high.
> >
> > --Mike
> >
> >
> > > Dear list,
> > >
> > > I was experimenting with MSI_MOSI_CMP_directory protocol with the 
> > > tester.
> > > With the little.trace on GEMS online documentation
> > > http://www.cs.wisc.edu/gems/doc/wiki/moin.cgi/How_do_I_understand_a_Protocol ,
> > > below is the protocol trace I got:
> > >
> > > Testing clear stats...Done.
> > > Reading trace from file 'little.trace'...
> > >       1   7  -1        Seq               Begin       >       [0x400, 
> > > line
> > > 0x400]
> > >       4   1   3    L1Cache                Load     NP>L1_IS  [0x400, 
> > > line
> > > 0x400]
> > >     141   1   0    L2Cache             L1_GETS  L2_NP>L2_IS  [0x400, 
> > > line
> > > 0x400]
> > >     390   0   0  Directory                GETS     NP>S      [0x400, 
> > > line
> > > 0x400]
> > >     635   1   0    L2Cache      Data_ext_ack_0  L2_IS>L2_SS  [0x400, 
> > > line
> > > 0x400]
> > >    1097   7  -1        Seq                Done       >       [0x400, 
> > > line
> > > 0x400] 1096 cycles NULL Yes
> > >    1097   1   3    L1Cache             L1_Data  L1_IS>L1_S   [0x400, 
> > > line
> > > 0x400]
> > >    1101   1  -1        Seq               Begin       >       [0x400, 
> > > line
> > > 0x400]
> > >    1104   0   1    L1Cache                Load     NP>L1_IS  [0x400, 
> > > line
> > > 0x400]
> > >    1139   0   0    L2Cache             L1_GETS  L2_NP>L2_IS  [0x400, 
> > > line
> > > 0x400]
> > >    1176   0   0  Directory                GETS      S>S      [0x400, 
> > > line
> > > 0x400]
> > >    1309   0   0    L2Cache      Data_ext_ack_0  L2_IS>L2_SS  [0x400, 
> > > line
> > > 0x400]
> > >    1445   1  -1        Seq                Done       >       [0x400, 
> > > line
> > > 0x400] 344 cycles NULL Yes
> > >    1445   0   1    L1Cache             L1_Data  L1_IS>L1_S   [0x400, 
> > > line
> > > 0x400]
> > >
> > > According to the documentation, the first column indicates the cycle. I
> > > don't understand why the operation cycles are so large. In my
> > > configuration,
> > >
> > > MEMORY_RESPONSE_LATENCY_MINUS_2: 78
> > > DIRECTORY_LATENCY: 80
> > > L2_RESPONSE_LATENCY: 6
> > > L1_RESPONSE_LATENCY: 3
> > > L1_REQUEST_LATENCY: 2
> > > L2_REQUEST_LATENCY: 4
> > > NETWORK_LINK_LATENCY: 40
> > >
> > > I don't understand why the cache operations would add up to, 1096 cycles
> > > for the first LD as an example. Could someone explain this please? By 
> > > the
> > > way, in the ruby config file, there is a TIMER_LATENCY: 10000. I wonder
> > > what this is.
> > >
> > > Thanks a lot! I appreciate your comments.
> > >
> > > Lei
>
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
> Use Google to search the GEMS Users mailing list by adding 
> "site:https://lists.cs.wisc.edu/archive/gems-users/"; to your search.