It's not really a bug, but it's definitely not the proper behavior.
The problem is that the CPI statistic is a formula, which is defined in
terms of two other stats: simTicks and committedInsts. simTicks is a
global stat of the number of total ticks that have been simulated,
which is why the CPI winds up being 100100 / 134. You'll need to keep
track of your own stat of the number of ticks simulated in the detailed
model, which can be defined as numCycles + idleCycles. I believe you
can define a statistic formula that is "numCycles + idleCycles", and
then use that formula in the CPI statistic formula (in place of
simTicks). In theory statistic formulas with other statistic formulas
in them should work, but I haven't tested that myself.
I've done a lot of work on getting full sampling to work on the older
version of m5, and this is one of the problems I ran into.
Unfortunately a lot of my work was with the Sampler, which has since
been removed, but I do have some outstanding changes to the CPU models
that I'll update once I get the chance.
Kevin
Vilas Sridharan wrote:
Hello all --
Thanks for the responses re: CPU switching; I seem to have gotten that
working.
I can now switch in the O3 CPU after some time running with the
AtomicSimpleCPU, and it seems to function properly. However, the CPI
stat in the output seems suspicious. When running with the
AtomicSimpleCPU for 100k cycles and the O3 CPU for 100 cycles, I get
the stats output attached at the end of this email. Note the CPI is
747, despite the fact that it appears to have committed 134
instructions in 100 cycles.
The reported CPI (747) is exactly (100100 / 134), or the total number
of simulated cycles (AtomicSimpleCPU + O3CPU) over the number of insts
committed by the O3 CPU.
Is there something I could be doing in my se.py file that would cause
this behavior, or is this a bug in the CPU model? My thought is the
latter - when I use the same se.py but switch in the AtomicSimpleCPU,
both the initial CPU and the switched-in CPU report the correct
numCycles value (100k and 100, respectively) - but I'm still not
familiar enough with python to know for sure.
Thanks in advance for the help!
-Vilas
m5stats.txt
-----
sim_insts
100135 # Number of instructions simulated
sim_ticks
100100 # Number of ticks simulated
system.cpu.idle_fraction
0 # Percentage of idle cycles
system.cpu.not_idle_fraction
1 # Percentage of non-idle cycles
system.cpu.numCycles
100001 # number of cpu cycles simulated
system.cpu.num_insts
100001 # Number of instructions executed
system.cpu.num_refs
49754 # Number of memory references
system.cpu.workload.PROG:num_syscalls
3 # Number of system calls
system.switch_cpu.commit.COM:branches
9 # Number of branches committed
system.switch_cpu.commit.COM:bw_lim_events
8 # number cycles where commit BW limit reached
system.switch_cpu.commit.COM:bw_limited
0 # number of insts not committed
due to BW limits
system.switch_cpu.commit.COM:committed_per_cycle.start_dist
# Number of insts commited each cycle
system.switch_cpu.commit.COM:committed_per_cycle.samples
88
system.switch_cpu.commit.COM:count
134 # Number of instructions committed
system.switch_cpu.commit.COM:loads
33 # Number of loads committed
system.switch_cpu.commit.COM:membars
0 # Number of memory barriers committed
system.switch_cpu.commit.COM:refs
66 # Number of memory references committed
system.switch_cpu.commit.COM:swp_count
0 # Number of s/w prefetches committed
system.switch_cpu.commit.branchMispredicts
7 # The number of times a branch was mispredicted
system.switch_cpu.commit.commitCommittedInsts
134 # The number of committed instructions
system.switch_cpu.commit.commitSquashedInsts
72 # The number of squashed insts skipped by
commit
system.switch_cpu.committedInsts
134 # Number of Instructions Simulated
system.switch_cpu.committedInsts_total
134 # Number of Instructions Simulated
system.switch_cpu.cpi
747.014925 # CPI: Cycles Per Instruction
system.switch_cpu.cpi_total
747.014925 # CPI: Total CPI of All Threads
_______________________________________________
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
