Nikos Anastopoulos wrote:
Hi,
I have some questions about m5 v2.0b1.
1) For SE mode, is there a PARMACS m5-specific implementation I could
use in order to write multithreaded (shared memory) programs? I know I
can use pthreads in FS mode instead, but I would like to exploit the
SMT feature which is currently available only in SE.
There is no PARMACS support in v2.0b1, since MP coherence isn't quite
working. v1.1 supports PARMACS for Tru64 via an m5-specific library.
I've been working on PARMACS/pthreads SE-mode support for v2.0 but
haven't finished it yet. The good news is that once I get it done it
will be (a) for Linux rather than Tru64 and (b) hopefully a little more
realistic than the Tru64 model, which relies too much on special m5
syscalls for things like condition variables.
In case that I can use the PARMACS model to write multithreaded apps, is
there a way then to map specific threads onto specific logical
processors (e.g. in a system with CMPs of SMTs, etc.)?
Not out of the box, but it should be possible to add.
2) How can I learn about all the available statistics that are
correlated with a specific component (e.g. caches or branch predictors)?
How can I get them printed? Are they printed on m5stats file by default,
or should I add some extra lines on the simulation script?
All meaningful (basically non-zero) stats are printed in the m5stats
file. It's my philosophy (that I've effectively imposed on everyone
else :-) ) that you never want to have to redo a multi-hour simulation
run just because you didn't collect some statistic that you didn't think
was relevant at the time you ran it but later changed your mind.
The easiest way to see all the stats for a particular object is to look
in the appropriate .cc file in that object's regStats() method. Looking
in the m5stats file is even easier, but you might miss some stats that
happened to be zero-valued for that run.
3) Which is the proposed way to get accurate measurements and statistics
for the execution of a specific app in FS mode (i.e., excluding all
statistics regarding the boot process and so on)? Should I add on my
.rcS file the "m5 dumpstats" command before and after the execution of
the application, and then get a diff?
There's a call to reset all the stats; I'm not sure if that happens
automatically when you do dumpstats or if it's a separate operation. Do
that before your app and then call dumpstats afterward and you'll get
the stats for your app directly.
4) How can I start m5 with a checkpoint obtained from a previous
execution? I am not sure I have read something about on the wiki pages...
Right now checkpointing only works in FS mode. I think there is
documentation on the wiki on how to use it.
Thanks in advance,
Nick.
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
