Hi Ricardo,
In some high-contention, coarse-grain transactional applications, like
Labyrinth, signatures may produce a positive effect in the execution time. In
that case, I guess that signatures generate false positives that stall
transactions, which serialize the execution, allowing the detection of
conflicts before they are produced. This fact may be good if the stalled
transaction aborts, because it must restore less lines. You can check this
behavior by looking at the time spent in abort recovery and the number of
conflicts.
LogTM-SE does not scale with Labyrinth. You can achieve more speed-up with other
configurations that reduce the number of aborts. However, check the speed-up
against a single-threaded execution. I think that it will not be 10x.
Finally, Labyrinth (and I think that sometimes other STAMP benchmarks) presents
some conflicts between Escape (stores in PRIV_MODE) and Store_XACT accesses.
These Escape accesses are treated as non-transactional stores in the coherence
protocol, which may produce inconsistency. I addressed this problem by changing
the coherence protocol and the conflict resolution policy to abort transactions
that conflict with these Escape Actions.
Best regards,
Marc Lupon
|
![[Computer Systems Lab]](http://www.cs.wisc.edu/pics/csl_logo.gif){kind=logo}