Hi Richard,

Checkpoints do NOT record cache state, so doing it the way you described will not yield a warm cache.  Recording cache state would require so much time and space it is not worth it.  Already, without cache state, checkpoint files are very large.

What you want to do is run linux from the beginning with just a simpleCPU (not a sampler), and dump a checkpoint at the point of your choice.  then, when you begin from the checkpoint, use the sampling CPU, which will run for a while in cacheCPU (to warmup caches) before switching to detailed.  does that make sense?

Finally, the m5 switchcpu instruction was never fully implemented.  but m5 i just a custom instruction we have, so that we can do little things that we want, like drop a checkpoint.  details can be found in m5/util/m5/m5.

lisa




On 4/13/06, Richard R. Zhang <[EMAIL PROTECTED] > wrote:
Hi Lisa,
Thanks a lot for your help. I have tried the sampler mode and it seems that the checkpoint works. But I have some
questions about the cache warm up. I change the default warm up cycles to 1e10, and start the linux from beginning.
The checkpoint is generated when the cache warm up command has finished. The following is part of my boot script,
which is modified from the original m5 maerts script.
     ...
        echo -n "waiting for server..."
        /usr/bin/netcat -c -l -p 8000
        BINARY=/benchmarks/netperf/netperf
        TEST="TCP_MAERTS"
        #SHORT_ARGS="-l -100k"
        SHORT_ARGS="-l 2"
        LONG_ARGS="-k16384,0 -K16384,0 -- -m 65536 -M 65536 -s 262144 -S 262144"

        SHORT="$BINARY -H $SERVER -t $TEST $SHORT_ARGS"
        LONG="$BINARY -H $SERVER -t $TEST $LONG_ARGS"

        echo "starting test..."
        echo "netperf warmup"
        echo $SHORT
        eval $SHORT

        /sbin/m5 ivlb 1
        /sbin/m5 resetstats
        /sbin/m5 dumpresetstats 200000000 2000000000
        /sbin/m5 checkpoint 1 0                 #  Q1
        /sbin/m5 checkpoint 200000000 2000000000
        echo "netperf benchmark"
        echo $LONG
        eval $LONG
        /sbin/m5 exit


The checkpoint generated by the command `/sbin/m5 checkpoint 1 0', shows that the curTick is a 10062394993, just a
little bigger than 1e10. My question is, when I restart the test from that checkpoint, has the cache warmed up? And
which mode will the m5 be at that time, the cacheCPU mode or detailedCPU mode?
BTW, would you pls give me a description about the user program `m5'? I tried to use the  `switchcpu' to switch the
cpu mode, but it seemed that the m5 simulator exited when execute that command. Thanks a lot!

Best wishes.

Richard
2006-04-13



发件人: Lisa Hsu
发送时间: 2006-04-11 22:24:19
收件人: zhangrui
抄送: m5sim-users
主题: Re: [m5sim-users] Checkpoint usage

m5 does not support checkpointing in detailed mode - there is just toooo much state to remember.  what we do is generate the checkpoint in simple CPU, and then beginning running again after the checkpoint with simple CPU or cacheCPU, and then transition to detailed CPU and take statistics from that point.

we usually use cacheCPU in order to warm up a cache system before going into detailed mode.  the number of cycles to do the warmup and the number of cycles to do the detailed run are defaulted to 1e9 and 200e6.  you can change these values if you wish.

lisa


On 4/11/06, zhangrui <[EMAIL PROTECTED] > wrote:
Hi all,
I have a question about the checkpoint usage.
When I use m5 to run netperf maerts test in a detailed full system mode with the default maerts scripts. I generat
a checkpoint to avoid Linux booting up process. But when I start m5 with -ECKPT_FILE=<checkpoint dir name> argument,
m5 reports a warning `warn: Not unserializing '': no section found in checkpoint.'. The messages in client console
show that the simulation process paused when jump to kernel. A few minute later, m5 reported `panic: PTE GH field != 0'
and exits.
Can m5 support the checkpoint unserialization in detailed full system mode? If so, how to configure it? I can't find
any document about this.

Thanks a lot.




Richard
2006-04-11



-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642
_______________________________________________
m5sim-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/m5sim-users


Reply via email to