Test Suite - Performance

From Libreswan
Jump to navigation Jump to search

Software - what is run by each test

Boot the VMs

Before a test can be run all the VMs are (re)booted. Consequently one obvious way to speed up testing is to reduce the amount of time it takes to boot:

  • make the boot faster - it should be around 1s
  • boot several machines in parallel - however booting is CPU intensive (see below for analysis)

To determine where a VM is spending its time during boot, use systemd-analyze blame (do several runs, the very first boot does extra configuration so is always be slower):

$ date ; ./testing/utils/kvmsh.py --boot cold l.east 'systemd-analyze time ; systemd-analyze critical-chain ; systemd-analyze blame'
Mon 10 Aug 2020 09:10:06 PM EDT
virsh 0.00: waiting 20 seconds for domain to shutdown
virsh 0.05: domain shutdown after 0.5 seconds
virsh 0.06: starting domain
virsh 11.07: got login prompt; sending 'root' and waiting 5 seconds for password (or shell) prompt
virsh 11.08: got password prompt after 0.1 seconds; sending 'swan' and waiting 5 seconds for shell prompt
virsh 12.00: we're in after 0.3 seconds!

[root@east ~]# systemd-analyze time ; systemd-analyze critical-chain ; systemd-analyze blame
Startup finished in 1.270s (kernel) + 1.837s (initrd) + 4.448s (userspace) = 7.557s 
multi-user.target reached after 4.411s in userspace
The time when unit became active or started is printed after the "@" character.
The time the unit took to start is printed after the "+" character.

multi-user.target @4.411s
└─sshd.service @4.356s +52ms
  └─network.target @4.350s
    └─systemd-networkd.service @1.236s +196ms
      └─systemd-udevd.service @1.003s +229ms
        └─systemd-tmpfiles-setup-dev.service @869ms +102ms
          └─kmod-static-nodes.service @644ms +115ms
            └─systemd-journald.socket
              └─system.slice
                └─-.slice
2.909s systemd-networkd-wait-online.service
 445ms systemd-udev-trigger.service        
 443ms systemd-vconsole-setup.service      
 229ms systemd-udevd.service               
 209ms systemd-journald.service            
 196ms systemd-networkd.service            
 180ms systemd-tmpfiles-setup.service      
 178ms systemd-logind.service              
 145ms auditd.service                      
 138ms source.mount                        
 124ms testing.mount                       
 115ms kmod-static-nodes.service           
 111ms systemd-journal-flush.service       
 110ms tmp.mount                           
 102ms systemd-tmpfiles-setup-dev.service  
  99ms systemd-modules-load.service        
  84ms systemd-remount-fs.service          
  71ms systemd-random-seed.service         
  58ms systemd-sysctl.service              
  57ms dbus-broker.service                 
  52ms sshd.service                        
  50ms systemd-userdbd.service             
  33ms systemd-user-sessions.service       
  24ms systemd-fsck-root.service           
  23ms dracut-shutdown.service             
  21ms systemd-update-utmp.service         
  13ms systemd-update-utmp-runlevel.service
   6ms sys-kernel-config.mount             
[root@east ~]# 

Run the Test Scripts

To establish a baseline, enumcheck-01, which pretty much nothing, takes ~2s to run the test scripts once things are booted:

w.runner enumcheck-01 32.08/32.05: start running scripts west:west.sh west:final.sh at 2018-10-24 22:00:44.706355
...
w.runner enumcheck-01 34.03/34.00: stop running scripts west:west.sh west:final.sh after 1.5 seconds

everything else is slower.

To get a list of script times:

$ awk '/: stop running scripts/ { print $3, $(NF-1) }' testing/pluto/*/OUTPUT/debug.log | sort -k2nr | head -5
newoe-05-hold-pass 295.6
newoe-04-pass-pass 226.7
ikev2-01-fallback-ikev1 212.5
newoe-10-expire-inactive-ike 205.6
ikev2-32-nat-rw-rekey 205.4

which can then be turned into a histogram:

Test-script-time-histogram.jpg

sleep

ping

timeout

Perform Post-mortem

This seems to be in the noise vis:

m1.runner ipsec-hostkey-ckaid-01 12:44:50.01: start post-mortem ipsec-hostkey-ckaid-01 (test 725 of 739) at 2018-10-25 09:40:49.748041
m1.runner ipsec-hostkey-ckaid-01 12:44:50.03: ****** ipsec-hostkey-ckaid-01 (test 725 of 739) passed ******
m1.runner ipsec-hostkey-ckaid-01 12:44:50.03: stop post-mortem ipsec-hostkey-ckaid-01 (test 725 of 739) after 0.2 seconds

KVM Hardware

What the test runs on

Disk I/O

Something goes here?

Memory

How much is needed?

CPU

Anything Here? Allowing use of HOST's h/w accelerators?

Docker Hardware

?

Tuning kvm performance

Internally kvmrunner.py has two work queues:

  • a pool of reboot threads; each thread reboots one domain at a time
  • a pool of test threads; each thread runs one test at a time using domains with a unique prefix

The test threads uses the reboot thread pool as follows:

  • get the next test
  • submit required domains to reboot pool
  • wait for domains to reboot
  • run test
  • repeat

By adjusting KVM_WORKERS and KVM_PREFIXES it is possible to:

  • speed up test runs
  • run independent testsuites in parallel

By adjusting KVM_LOCALDIR it is possible to:

  • use a faster disk or even tmpfs (on /tmp)


KVM_WORKERS=... -- the number of test domains (machines) booted in parallel

Booting the domains is the most CPU intensive part of running a test, and trying to perform too many reboots in parallel will bog down the machine to the point where tests time out and interactive performance becomes hopeless. For this reason a pre-sized pool of reboot threads is used to reboot domains:

  • the default is 1 reboot thread limiting things to one domain reboot at a time
  • KVM_WORKERS specifies the number of reboot threads, and hence, the reboot parallelism
  • increasing this allows more domains to be rebooted in parallel
  • however, increasing this consumes more CPU resources

To increase the size of the reboot thread pool set KVM_WORKERS. For instance:

$ grep KVM_WORKERS Makefile.inc.local
KVM_WORKERS=2
$ make kvm-install kvm-test
[...]
runner 0.019: using a pool of 2 worker threads to reboot domains
[...]
runner basic-pluto-01 0.647/0.601: 0 shutdown/reboot jobs ahead of us in the queue
runner basic-pluto-01 0.647/0.601: submitting shutdown jobs for unused domains: road nic north
runner basic-pluto-01 0.653/0.607: submitting boot-and-login jobs for test domains: east west
runner basic-pluto-01 0.654/0.608: submitted 5 jobs; currently 3 jobs pending
[...]
runner basic-pluto-01 28.585/28.539: domains started after 28 seconds

Only if your machine has lots of cores should you consider adjusting this in Makefile.inc.local.


KVM_PREFIXES=... -- create a pool of test domains (machines)

Tests spend a lot of their time waiting for timeouts or slow tasks to complete. So that tests can be run in parallel the KVM_PREFIX provides a list of prefixes to add to the host names forming unique domain groups that can each be used to run tests:

  • the default is no prefix limiting things to a single global domain pool
  • KVM_PREFIXES specifies the domain prefixes to use, and hence, the test parallelism
  • increasing this allows more tests to be run in parallel
  • however, increasing this consumes more memory and context switch resources

For instance, setting KVM_PREFIXES in Makefile.inc.local to specify a unique set of domains for this directory:

$ grep KVM_PREFIX Makefile.inc.local
KVM_PREFIX=a.
$ make kvm-install
[...]
$ make kvm-test
[...]
runner 0.018: using the serial test processor and domain prefix 'a.'
[...]
a.runner basic-pluto-01 0.574: submitting boot-and-login jobs for test domains: a.west a.east

And setting KVM_PREFIXES in Makefile.inc.local to specify two prefixes and, consequently, run two tests in parallel:

$ grep KVM_PREFIX Makefile.inc.local
KVM_PREFIX=a. b.
$ make kvm-install
[...]
$ make kvm-test
[...]
runner 0.019: using the parallel test processor and domain prefixes ['a.', 'b.']
[...]
b.runner basic-pluto-02 0.632/0.596: submitting boot-and-login jobs for test domains: b.west b.east
[...]
a.runner basic-pluto-01 0.769/0.731: submitting boot-and-login jobs for test domains: a.west a.east

creates and uses two dedicated domain/network groups (a.east ..., and b.east ...).

Finally, to get rid of all the domains use:

$ make kvm-uninstall

or even:

$ make KVM_PREFIX=b. kvm-uninstall

Two domain groups (e.x., KVM_PREFIX=a. b.) seems to give the best results.

Note that this is still somewhat experimental and has limitations:

  • stopping parallel tests requires multiple control-c's
  • since the duplicate domains have the same IP address, things like "ssh east" don't apply; use "make kvmsh-<prefix><domain>" or "sudo virsh console <prefix><domain" or "./testing/utils/kvmsh.py <prefix><domain>".


KVM_LOCALDIR=/tmp/pool -- the directory containing the test domain (machine) disks

To reduce disk I/O, it is possible to store the test domain disks in ram using tmpfs and /tmp. Here's a nice graph illustrating what happens when the option is set:

Diskstats iops-day.png


Recommendations

Some Analysis

The test system:

  • 4-core 64-bit intel
  • plenty of ram
  • the file mk/perf.sh

Increasing the number of parallel tests, for a given number of reboot threads:

Tests-vs-reboots.png

  • having #cores/2 reboot threads has the greatest impact
  • having more than #cores reboot threads seems to slow things down

Increasing the number of reboots, for a given number of test threads:

Reboots-vs-tests.png

  • adding a second test thread has a far greater impact than adding a second reboot thread - contrast top lines
  • adding a third and even fourth test thread - i.e., up to #cores - still improves things

Finally here's some ASCII art showing what happens to the failure rate when the KVM_PREFIX is set so big that the reboot thread pool is kept 100% busy:

                  Fails  Reboots  Time
     ************  127      1     6:35  ****************************************
   **************  135      2     3:33  *********************
  ***************  151      3     3:12  *******************
  ***************  154      4     3:01  ******************

Notice how having more than #cores/2 KVM_WORKERS (here 2) has little benefit and failures edge upwards.

Desktop Development Directory

  • reduce build/install time - use only one prefix
  • reduce single-test time - boot domains in parallel
  • use the non-prefix domains east et.al. so it is easy to access the test domains using tools like ssh

Lets assume 4 cores:

KVM_WORKERS=2
KVM_PREFIX=''

You could also add a second prefix vis:

KVM_PREFIX= '' a.

but that, unfortunately, slows down the the build/install time.

Desktop Baseline Directory

  • do not overload the desktop - reduce CPU load by booting sequentially
  • reduce total testsuite time - run tests in parallel
  • keep separate to development directory above

Lets assume 4 cores

  • KVM_WORKERS=1
  • KVM_PREFIX= b1. b2.

Dedicated Test Server

  • minimize total testsuite time
  • maximize CPU use
  • assume only testsuite running

Assuming 4 cores:

* KVM_WORKERS=2
* KVM_PREFIX= '' t1. t2. t3.