The incorrect resuts

[pocean]

Dear code authors:
I’m using the OptaneStudy, but I have encountered some problems in using it.
I use it in Fedora 27 and linux kernel 4.13.16. I config the Octane use these commands:
1. Create pmem devices
ipmctl create -goal PersistentMemoryType=AppDirect
ndctl create-namespace
and get these regions:
[
{
"dev":"namespace3.0”,
"mode":"fsdax”,
"map":"dev”,
"size":133175443456,
"uuid":"8824fbe5-6b52-4463-97b3-ee1c76cdf685”,
"blockdev":"pmem3”
},
{
"dev":"namespace2.0”,
"mode":"fsdax”,
"map":"dev”,
"size":133175443456,
"uuid":"904229c9-0cc0-425e-a288-c90cfd9a1ba8”,
"blockdev":"pmem2”
}
]

2. Execute the mount.sh
sh /root/OptaneStudy/src/testscript/mount.sh /dev/pmem2 /dev/pmem3

3. Test Latency
echo task=1,op=0 > /proc/lattester

4. See the output using dmesg
dmesg | tail -n 100
and get the output like this:
[ 266.020088] {0}[0]load-fence-64 avg 2729, stddev^2 21261321, max 58924, min 864 
[ 266.064615] {0}Running ntload-fence-64 
[ 268.347241] {0}[1]ntload-fence-64 avg 2728, stddev^2 2328676, max 22524, min 866 
[ 268.391851] {0}Running store-fence-64 
[ 269.926788] {0}[2]store-fence-64 avg 18446744073709550438, stddev^2 5080516, max 58952, min 18446744069414592274 
[ 269.994989] {0}Running store-clflush-64 
[ 270.868930] {0}[3]store-clflush-64 avg 18446744073709548557, stddev^2 19686969, max 53560, min 18446744069414585292

I'm afraid this result is incorrect. May I ask what is wrong with my testing process.

[sheepx86]

The rdtscp instruction (read TSC) is not a serializing instruction. There are possibilities that two rdtscp (with mfence) end up in a negative number. There are a few ways to mitigate this issue, since we are trying to measure "memory latency", some of the methods may affect the result.

What we did was simply disregarding the outliers. See the parsing code at:
src/testscript/parsing/10_parse_basic.py

Reference:
https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf

[pocean]

@sheepx86 Thanks for your replying. I disregard the outliers and got the results:
[ 203.766288] {0}Running load-fence-64
[ 206.145701] {0}[0]load-fence-64 avg 1620, stddev^2 37405456, max 9908, min 1000
[ 206.190273] {0}Running ntload-fence-64
[ 208.478466] {0}[1]ntload-fence-64 avg 2702, stddev^2 1695204, max 9988, min 868
My cpu is : Intel(R) Xeon(R) Gold 6246 CPU @ 3.30GHz, so I think the latency of "load-fence-64" is 1000/3.3=303ns. However, according to the paper, the latency of sequential Optane load is ~169ns. In addition, the random Optane read latency I tested have similar results, it seems that there is no buffer on the Optane.

[TheNetAdmin]

Hi, have you tried:

1. Set CPU to 'performance mode', like here:

https://github.com/TheNetAdmin/LENS/blob/663e3f367563e3927edae9a1ee2e840a78977a29/scripts/utils/launch.sh#L33-L36

2. Turn off cache prefetcher, like here:

https://github.com/TheNetAdmin/LENS/blob/663e3f367563e3927edae9a1ee2e840a78977a29/scripts/utils/launch.sh#L39

[yzouchen]

My situation is totally the same, and I already tried Set CPU to 'performance mode and Turn off cache prefetcher.
Cant get a similar result of task 1. Any suggestions?