tests: introduce a framework for testing migration performance

This introduces a moderately general purpose framework for
testing performance of migration.

The initial guest workload is provided by the included 'stress'
program, which is configured to spawn one thread per guest CPU
and run a maximally memory intensive workload. It will loop
over GB of memory, xor'ing each byte with data from a 4k array
of random bytes. This ensures heavy read and write load across
all of guest memory to stress the migration performance. While
running the 'stress' program will record how long it takes to
xor each GB of memory and print this data for later reporting.

The test engine will spawn a pair of QEMU processes, either on
the same host, or with the target on a remote host via ssh,
using the host kernel and a custom initrd built with 'stress'
as the /init binary. Kernel command line args are set to ensure
a fast kernel boot time (< 1 second) between launching QEMU and
the stress program starting execution.

None the less, the test engine will initially wait N seconds for
the guest workload to stablize, before starting the migration
operation. When migration is running, the engine will use pause,
post-copy, autoconverge, xbzrle compression and multithread
compression features, as well as downtime & bandwidth tuning
to encourage completion. If migration completes, the test engine
will wait N seconds again for the guest workooad to stablize on
the target host. If migration does not complete after a preset
number of iterations, it will be aborted.

While the QEMU process is running on the source host, the test
engine will sample the host CPU usage of QEMU as a whole, and
each vCPU thread. While migration is running, it will record
all the stats reported by 'query-migration'. Finally, it will
capture the output of the stress program running in the guest.

All the data produced from a single test execution is recorded
in a structured JSON file. A separate program is then able to
create interactive charts using the "plotly" python + javascript
libraries, showing the characteristics of the migration.

The data output provides visualization of the effect on guest
vCPU workloads from the migration process, the corresponding
vCPU utilization on the host, and the overall CPU hit from
QEMU on the host. This is correlated from statistics from the
migration process, such as downtime, vCPU throttling and iteration
number.

While the tests can be run individually with arbitrary parameters,
there is also a facility for producing batch reports for a number
of pre-defined scenarios / comparisons, in order to be able to
get standardized results across different hardware configurations
(eg TCP vs RDMA, or comparing different VCPU counts / memory
sizes, etc).

To use this, first you must build the initrd image

 $ make tests/migration/initrd-stress.img

To run a a one-shot test with all default parameters

 $ ./tests/migration/guestperf.py > result.json

This has many command line args for varying its behaviour.
For example, to increase the RAM size and CPU count and
bind it to specific host NUMA nodes

 $ ./tests/migration/guestperf.py \
       --mem 4 --cpus 2 \
       --src-mem-bind 0 --src-cpu-bind 0,1 \
       --dst-mem-bind 1 --dst-cpu-bind 2,3 \
       > result.json

Using mem + cpu binding is strongly recommended on NUMA
machines, otherwise the guest performance results will
vary wildly between runs of the test due to lucky/unlucky
NUMA placement, making sensible data analysis impossible.

To make it run across separate hosts:

 $ ./tests/migration/guestperf.py \
       --dst-host somehostname > result.json

To request that post-copy is enabled, with switchover
after 5 iterations

 $ ./tests/migration/guestperf.py \
       --post-copy --post-copy-iters 5 > result.json

Once a result.json file is created, a graph of the data
can be generated, showing guest workload performance per
thread and the migration iteration points:

 $ ./tests/migration/guestperf-plot.py --output result.html \
        --migration-iters --split-guest-cpu result.json

To further include host vCPU utilization and overall QEMU
utilization

 $ ./tests/migration/guestperf-plot.py --output result.html \
        --migration-iters --split-guest-cpu \
	--qemu-cpu --vcpu-cpu result.json

NB, the 'guestperf-plot.py' command requires that you have
the plotly python library installed. eg you must do

 $ pip install --user  plotly

Viewing the result.html file requires that you have the
plotly.min.js file in the same directory as the HTML
output. This js file is installed as part of the plotly
python library, so can be found in

  $HOME/.local/lib/python2.7/site-packages/plotly/offline/plotly.min.js

The guestperf-plot.py program can accept multiple json files
to plot, enabling results from different configurations to
be compared.

Finally, to run the entire standardized set of comparisons

  $ ./tests/migration/guestperf-batch.py \
       --dst-host somehost \
       --mem 4 --cpus 2 \
       --src-mem-bind 0 --src-cpu-bind 0,1 \
       --dst-mem-bind 1 --dst-cpu-bind 2,3
       --output tcp-somehost-4gb-2cpu

will store JSON files from all scenarios in the directory
named tcp-somehost-4gb-2cpu

Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
Message-Id: <1469020993-29426-7-git-send-email-berrange@redhat.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>

1 files changed, 2 insertions, 0 deletions

diff --git a/configure b/configure
index 5ada56d..6ffa4a8 100755
--- a/configure
+++ b/configure
@@ -3121,6 +3121,7 @@ else
       if test "$found" = "no"; then
         LIBS="$pthread_lib $LIBS"
       fi
+      PTHREAD_LIB="$pthread_lib"
       break
     fi
   done
@@ -5537,6 +5538,7 @@ echo "LDFLAGS_NOPIE=$LDFLAGS_NOPIE" >> $config_host_mak
 echo "LD_REL_FLAGS=$LD_REL_FLAGS" >> $config_host_mak
 echo "LIBS+=$LIBS" >> $config_host_mak
 echo "LIBS_TOOLS+=$libs_tools" >> $config_host_mak
+echo "PTHREAD_LIB=$PTHREAD_LIB" >> $config_host_mak
 echo "EXESUF=$EXESUF" >> $config_host_mak
 echo "DSOSUF=$DSOSUF" >> $config_host_mak
 echo "LDFLAGS_SHARED=$LDFLAGS_SHARED" >> $config_host_mak