[Gems-users] Tourmaline issues with running microbenchmark tm-deque

["Shougata Ghosh (shougata@xxxxxxxxxxxxx)" <shougata@xxxxxxxxxxxxx>]

Hi
I'm having problems running the tm-deque microbenchmark with Tourmaline. I modified the tm-deque microbenchmark by adding a barrier so that all the parallel threads start running at the same time. I wanted to model high contention among the transactional threads. When I run tm-deque with 10 threads and 30 operations, Tourmaline works fine.
However, when I remove the barrier and then run the microbenchmark again with 10 threads and 30 operations, the simulated machine crashes! With no barriers and 10 threads and 10 operations, tm-deque runs ok.

Here's my complete setup:
-----------------------------------------------------
Simics version 2.2.19
GEMS version 1.3

16 UltraSPARC-III+ processors running Solaris 10.

In the microbenchmark, I bind each thread to a particular processor so that they are not context switched during a transaction. Also, I bind the entire microbenchmark to a processor set so that it's not conflicting with other running processes.

Tourmaline setup:
Initializing TourmalineConfig... Done.
Creating Transaction Controller:
Controller_type: Generic
number_of_procs: 16
Creating Profiler... Done.

Tourmaline Module Initialized.
Tourmaline Configuration
------------------
simics_version: simics-2.2.19

g_RANDOM_SEED: 1
SPECIFIED_TRANSACTION_CONTROLLER: Generic
g_NUMBER_OF_PROCESSORS: 16
g_NUMBER_OF_ALLOWED_ABORTS: 10
g_SERIALIZATION_WARNING_ENABLE: true
g_PRINT_XACT_STATUS: true
g_SERIALIZER_N_TRANSACTIONS: 0

-------------------------------------------------------

The script used to run the microbenchmark:
#Read in a configuration

#Disable STCs here
dstc-disable

#Disable breakpoint before sync-ing processors
magic-break-disable

#Use 1 cycle cpu-switch-time
cpu-switch-time (1)
c 10000

#Print new cpu switch time
cpu-switch-time

magic-break-enable

load-module tourmaline
tourmaline0.init

run-python-file filename = setup_hap_breakpoint.sh
run-command-file file = setup_run_benchmark.sh

<At the breakpoint, I remove the magic-breakpoint callback hap and load the simcis tracer module>

Here's the output I get in the simulated processor window after tm-deque has run for sometime:

panic[cpu5]/thread=300023c5300: bad kernel MMU trap at TL 2

%tl %tpc              %tnpc             %tstate           %tt
 1  0000000000012440  0000000000012444  4482041601        04b
    %ccr: 44  %asi: 82  %cwp: 1  %pstate: 416<MG,PEF,PRIV,IE>
 2  000000000100733c  0000000001007340  9182001502        034
    %ccr: 91  %asi: 82  %cwp: 2  %pstate: 15<PEF,PRIV,AG>
%g0-3: 0000000000000000 0000000000000005 0000000005ca2b00 0000000005ca2b00
%g4-7: 000000000000000f 0000000000000034 0000000000000000 0000000000000002
Register window 2, caller 12300
%o0-3: 00000000ff36ba4c 0000000000000610 000000000000000b 0000000000000000
%o4-7: 0000000000000000 0000000084b69543 00000000ff1f3de0 00000000ff36b9d8
%l0-3: 0000000000012440 0000000000012444 0000004482041601 000000000100ae5c
%l4-7: 00000000ff36ba40 00000000ff36ba3c 0000000000000000 00000000ff1f468f
%i0-3: 000000000000005d 0000000000000010 0000000000000004 00000000000002a1
%i4-7: 0000000000000000 00000000425b4aa1 00000000ff1f3f30 0000000000012300
Register window 1, caller ff33fff0
%l0-3: 0000000000000010 0000000000000002 0000000000023a18 0000000000033800
%l4-7: 0000000000033800 0000000000023400 0000000000004001 00000000000fc000
%i0-3: 0000000000000003 00000000ff1fc000 0000000000000000 0000000000000000
%i4-7: 00000000000121e0 0000000000000000 00000000ff1fbfa0 00000000ff33fff0

I'm quite sure the problem is with Tourmaline because it runs fine if I don't load Tourmaline.
Has anyone got any idea what's causing this panic? I'd really appreciate any response.
Thanks a lot
-shougata