James,
Thanks for the post. Everything you said looks correct to me. Your
email can now serve as the official documentation on the matter :-).
Note that the phase0_cpus and phase1_cpus vectors can have multiple CPUs
to allow you to do sampling on a multiprocessor or multi-system
configuration. It's also worth repeating (from your earlier mail to the
list) that the statistics get reset when you switch, so the final stats
only reflect that last phase. (Somebody chime in if you have anything
else to add.)
Right now our only use of the sampler is for a one-time transition from
a SimpleCPU-based warm-up phase to a FullCPU-based measurement phase.
The key problem with switching in the other direction is making sure
that the CPU and the memory system have reached a steady state (i.e.
with no outstanding events that will cause trouble when they get
processed after the switchover). We have code in there that takes care
of the memory-system side of things (basically stalling the CPU until
all outstanding memory transactions complete before transitioning to the
non-event mode), but haven't worked on the CPU side of the problem (not
that it's necessarily that hard, we just haven't had the need to work on
it). If you're interested in doing continuous sampling, it's probably
not a huge amount of work to make that happen.
Steve
James Srinivasan wrote:
I'm trying to use the Sampler module to switch between functional
and performance simulation using phase analysis approaches like
SimPoint (UCSD). I am trying to us the Sampler module in ALPHA_SE
mode, but am not sure how to write the configuration files.
First off I'm not sure this is supported with the current m5 release.
sampler.cc:167 says:
//We can't switch back from detailed yet, just finish sampling
//Temporary stop-gap
when trying to switch from phase1 back to phase0.
Since sampling seems to have resulted in some confusion on the list I'll
write up what I've managed to work out so far. To use the sampler you
need to instantiate it like:
root.sampler = Sampler(phase0_cpus = [root.cpu0], phase1_cpus =
[root.cpu1], periods = [1e7, 2e7])
in your config file. The first period tells you how long phase0 will
last, the second period is how long phase1 will last. The intention is
to revert back to phase0 after phase1 but right now the code just exits
simulation cleanly. The critical bit you need to do is ensure the two
CPU models are hooked up to the same memory system and are running the
same workload. The way I'm doing this is by instantiating the memory
items and the workload in the root object as follows:
root=Root()
root.toL2Bus = ToL2Bus()
root.toMemBus = ToMemBus()
root.dcache = DL1(out_bus=Parent.toL2Bus)
root.icache = IL1(out_bus=Parent.toL2Bus)
root.l2 = L2(in_bus=Parent.toL2Bus, out_bus=Parent.toMemBus)
root.ram = SDRAM(in_bus=Parent.toMemBus)
I then instantiate the CPU models as:
root.cpu0 = MySimpleCPU()
root.cpu1 = MyDetailedCPU()
Inside the definition of MySimpleCPU and MyDetailedCPU (based on
SimpleCPU and DetailedCPU respectively) the icache and
dcache attributes are set to Parent.icache and Parent.dcache
I then instantiate the sampler as above. Finally I instantiate the
workload:
bm = Benchmarks.GzipSource()
root.cpu0.workload = bm
root.cpu1.workload = bm
I'm not 100% convinced this is the correct way and I might have
forgotten some details but from some quick simulations this morning it
seems ok. Can post my full config files if you like,
Hope this helps,
James
-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems? Stop! Download the new AJAX search engine that makes
searching your log files as easy as surfing the web. DOWNLOAD SPLUNK!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=103432&bid=230486&dat=121642
_______________________________________________
m5sim-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/m5sim-users
