Explore performance with genwqe_mt_perf

To get an idea how the hardware accelerator performs on your system you can of course use the accelerated version of gzip/gunzip. But if you like to see what you get in a multithreaded environment, you can use zlib_mt_perf and the associated genwqe_mt_perf script.

Install the tools with:

sudo make install

You can toggle between hardware and software operation mode. And you can see the influence of running multiple threads on the accumulated throughput of the solution.

Here an example using software compression/decompression:

~/genwqe-user$ genwqe_mt_perf -ASW

And here is how to start the measurement for the GenWQE accelerator card 0:

~/genwqe-user$ genwqe_mt_perf -AGENWQE -C0

Note that you can either use the auto-generated test-data in /tmp/test_data.bin or that you can write your one /tmp/test_data.bin. We are mostly using a Linux kernel tar (see www.kernel.org) file, this gives more realistic throughput values than random data, which does not compress at all. Using the 1st 32MiB of the kernel.tar file gives you shorter runtime.

To measure the CAPI version of the card, use:

~/genwqe-user$ genwqe_mt_perf -ACAPI -C0

If you have more than one card, try setting the card parameter to -1 (-C -1). That allows the library to automatically distribute the load across the available cards. This should show a positive effect once the load is high enough.

To get measurements without much influence of I/O, have /tmp to be a tmpfs for example:

sudo mount -t tmpfs /tmp /tmp

Some installations use already a ram filesystem for /tmp, some do not. Do not forget to umount if you have mounted it as tmpfs.

If you like to use the -l feature to monitor the CPU load during the experiment, you need to install the following packages:

apt-get install gnuplot
apt-get install sysstat

Enable the sysstat service:

open "/etc/default/sysstat" , Change ENABLED="true"

And start it:

service sysstat restart