[Gems-users] Simics abort when running transactional microbenchmark

[Yunyun Jiang <jiangyunyuntt@xxxxxxxxx>]

Hi, all,

I inserted some MAGIC_BREAKPOINT in the following functions in "$GEMS/microbenchmark/transactional/common/transaction.c":

void compensation_handler()

void commit_handler()

void register_compensation_handler()

void register_commit_handler()

void register_compensating_action()

void register_commit_action()

void tm_unroll_log_entry()

void tm_execute_compensation()

void tm_execute_commit_action()

void tm_execute_commit_actions()

void tm_release_isolation()

void tm_execute_summary_signature_conflict_handlers()

void tm_register_summary_signature_conflict_handler()

void tm_init_summary_signature_conflict_handlers()

void tm_log_unroll()

void tm_trap_handler()

void transaction_manager_stub()

void touch_log()

void walk_log()

void init_log()

void set_log_base()

void set_handler_address()

void set_transaction_registers()

void ruby_watch()

Then I ran the "paging" microbenchmark in "$GEMS/microbenchmark/transactional/paging" with "./paging 2 2" on the 4-core simulator. I have got a few break points at first:

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

[cpu0] v:0x0000000000013118 p:0x0003bf03118  magic (sethi 0x40000, %g0)

set_handler_address () in paging

simics> c

 1955659 0 [0,0] Setting Handler Address: [0x12bf0, line 0x12bc0]

......

[cpu1] v:0x0000000000013118 p:0x0003bf03118  magic (sethi 0x40000, %g0)

Setting new inspection cpu: cpu1

set_handler_address () in paging

simics> c

 1957286 1 [1,0] Setting Handler Address: [0x12bf0, line 0x12bc0]

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

But the simulation was aborted soon and showing:

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

1994032 1 SHOULD TRAP IN NON_XACT 0 *****************WARNING********************

simics-common: log_tm/TransactionSimicsProcessor.C:595: void TransactionSimicsProcessor::retiredMemRef(memory_transaction_t*): Assertion `0' failed.

Abort (SIGABRT) in main thread

Crash stack trace:

#0 0x00002ab85db86765 <unknown>

...

#102 0x00002ab85e59cba0 <unknown>

The simulation state has been corrupted. Simulation cannot continue.

Please restart Simics.

[cpu1] v:0x00000000000117b8 p:0x00039e917b8  stw %g0, [%g1 + %g0]

slaveLoop () in paging

simics> 

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

Besides, the "paging" benchmark will continue running even for a couple of days if there are no MAGIC_BREAKPOINT inserted in trancaction.c. Is that normal?

Could any one please tell me where am I doing wrong and where is the transaction conflict handler of LogTM in transaction.c or somewhere else? Thanks!

Best regards, 

Yunyun

-- 
JIANG Yunyun
Institute of High Performance Computing,
Department of Computer Science and Technology,
Tsinghua University, Beijing, China, 100084.