2. • Things to know about Global Task Scheduling (GTS).
• MP patchset description and how the solution works.
• Configuration parameters at various levels.
• Continuous integration at Linaro.
Today’s Presentation:
3. • This presentation is the lighter version of two
presentation Linaro has on GTS.
• The other runs for about 75 minutes and goes much
deeper in the solution.
• If you are interested in the in-depth version please
contact Joe Bates: joe.bates@linaro.org
Other Presentations on GTS:
4. • A set of patches enacting Global Task Scheduling(GTS).
• Developed by ARM Ltd.
• GTS modifies the Linux scheduler in order to place tasks
on the best possible CPU.
• Advantages:
• Take full advantage of the asynchronous nature of b.L architecture.
• Maximum performance
• Minimum power consumption
• Better benchmark scores for thread-intensive benchmarks.
• Increased responsiveness by spinning off new tasks on big CPUs.
• Decreases power consumption, specifically with small-task packing.
What is the MP Patchset?
5. • In a tarball from the release page:
• Always look for the latest “vexpress-lsk” release on release.linaro.org
- ex. for January:
http://releases.linaro.org/14.01/android/vexpress-lsk
• February should look like:
http://releases.linaro.org/14.02/android/vexpress-lsk
• In the Linaro Stable Kernel:
https://git.linaro.org/gitweb?p=kernel/linux-linaro-stable.git;a=summary
Where to get it
6. • In the ARM big LITTLE MP tree:
https://git.linaro.org/gitweb?p=arm/big.LITTLE/mp.git;a=summary
** Linaro doesn’t rebase the MP patchset on other kernels
than the Linaro Stable Kernel.
Where to get it (continued)
7. • General Overview:
• The Linux kernel builds a hierarchy of scheduling domains at boot
time. The order is (Linux convention):
• Sibling (for Hyperthreading)
• MC - multi-core
• CPU - between clusters
• NUMA
• To understand how the kernel does this:
• Enable CONFIG_SCHED_DEBUG and
• set “sched_debug=1” on the kernel cmd line
• In a pure SMP context load balancing is done by
spreading tasks evenly among all processors.
• Maximisation of CPU resources
• Run-to-completion model
MP Patchset Description
9. • Classic load balancing between CPU domains (i.e big
and LITTLE) is disabled.
• A derivative of Paul Turner’s “load_avg_contrib” metric is
used to decide if a task should be moved to another
HMP domain.
Paul’s work: http://lwn.net/Articles/513135/
• Migration of tasks among the CPU domains is done by
comparing their loads with migration thresholds.
• By default, all new user tasks are placed on the big
cluster.
How MP Works
11. Load Average Contribution and Decay
Plotting of the “runnable_avg_sum” metric introduced by Paul
Turner
12. • Paul Turner introduced the load average contribution
metric in his work on per-entity load tracking:
load_avg_contrib = task->weight * runnable_average
where runnable_average is:
runnable_average = runnable_avg_sum / runnable_avg_period
• runnable_avg_sum and runnable_avg_period are
geometric series.
• load_avg_contrib is good for scheduling decisions but
bad for task migration i.e, weight scaling doesn’t reflect
the true time spent by a task in the runnable state.
Per Entity Load Tracking
13. • The MP patchset introduces the load average ratio:
load_avg_ratio = NICE_0_LOAD * runnable_average
• The load average ratio allows for the comparison of
tasks without their weight factor, giving the same
perspective for all of them.
• At migration time the load average ratio is compared
against two thresholds:
• hmp_up_threashold
• hmp_down_threashold
Load Average Ratio
14. UP and Down Migration thresholds
A task’s load is compared to the up and down
migration threshold during the MP domain
balancing process.
* Source: ARM Ltd.
15. • The Linux scheduler will separate CPUs into domains.
• Tasks are spread out among the domains as equally as
possible.
• For GTS load balancing at the CPU domain level is
disabled.
• GTS will move tasks between CPU domains using a
derivative of the load average contribution and a couple
of thresholds.
• But when is GTS moving tasks between the CPU
domains?
What We’ve Learned So Far
16. • 4 task migration points:
• When tasks are created (fork migration).
• At wakeup time (wakeup migration).
• With every scheduler tick (forced migration).
• When a CPU is about to become idle (idle pull).
Task Migration Points
17. • When tasks are created (fork migration):
• Done by setting the task’s load statistics to their maximum value.
• Tasks are placed on big CPUs unless they are:
• Kernel Threads
• Forked from init i.e, Android services.
• Android apps are forked from Zygote, hence go on big CPUs.
• Tasks are eventually migrated down if they aren’t heavy enough.
Fork Migration
18. • At wakeup time (wakeup migration):
• When a task is to be placed on a CPU, the scheduler will normally
prefer:
• The previous CPU the task ran on
• Or one in the same package.
• For GTS, the decision is based on the load a task had before it was
suspended:
• if load(task) > hmp_up_threshold, select more potent HMP domain
• if load(task) < hmp_down_threshold, select less powerful HMP
domain
• What happened in the past is likely to happen again.
Wakeup Migration
19. • With every scheduler tick (forced migration):
• Every CPU in the system has a scheduler tick.
• With each tick (minimum interval of 1 jiffies) a CPU’s runqueue is
rebalanced if event due.
• Each time the load balancer runs, the MP code will inspect the
runqueue of all CPUs in the system:
• If LITTLE CPU → can a task be moved to big cluster?
• if ((big CPU ) && (CPU overloaded)) → offload lightest task.
• When offloading, always select an idle CPU to ensure CPU availability
for the task.
• So that tasks can be migrated as quickly as possible as domains can
stay balanced for a long time.
Forced Migration
20. • When a CPU is about to become idle(idle pull):
• When a CPU is about to go idle the scheduler will attempt to pull
tasks away from other CPUs in the same domain.
• Happens only if the CPU average idle time is more than the estimated
migration cost.
• Balancing within a domain is left to normal scheduler operation.
• If the scheduler didn’t find any task to pull and CPU is in big cluster:
• Go through the runqueues of all online CPUs in the LITTLE cluster.
• If a task’s load is above threshold, move it to a CPU in the big cluster.
• When moving a task, always look for the least loaded CPU.
Idle Pull
22. • Scheduler will try to fit as many small task on a single
CPU as possible.
• A small task is =< 90% of NICE_0_LOAD, i,e 921
• Done on the LITTLE cluster only to make sure tasks on
the big cluster have all the CPU time they need.
• Takes place when a task is waking up:
• Using the tracked load of CPU runqueues and tasks.
• Saturation threshold to make sure tasks offloaded from
the big domain can keep being serviced.
• Effects of enabling small task packing:
• CPU operating point may increase → CPUfreq governor will kick in.
• Wakeup latency of task may increase → more tasks to run.
Small Task Packing
23. • Load balancing at the CPU domain level is disabled to
favour the GTS scheme.
• GTS works by comparing a task’s runnable load ratio
and migrating it to a different HMP domain if need be.
• There are 4 migration points:
• At creation time.
• At wakeup time.
• Every rebalance.
• When a CPU is about to become idle.
• Small task packing when CPU gating is possible.
Key Things to Remember
24. • GTS doesn’t hotplug CPUs and is not concerned at all
with hotplugging
• When hotplugging:
• It takes too long to bring a CPU in and out of service
• All smpboot threads need to be stopped.
• “stop_machine” threads suspend interrupts on all online CPUs.
• IRQs on the swapped CPU are diverted to another CPU.
• All processes in swapped CPU’s runqueue are migrated.
• CPU is taken out of coherency.
• More CPUs means longer hotplug time per CPU.
• Very expensive to make a CPU coherent with the domain hierarchy
again.
• The system needs intelligence to determine when CPUs will be
swapped in and out.
One Last Remark
25. • The GTS solution itself has a number of parameters that
can be tuned. Examples:
• From /sys/kernel/hmp:
• up_threshold, down_threshold for task migration limits
• load_avg_period_ms and frequency_invariant_load_scale
• From the code:
• runqueue saturation when doing small task packing
• Amount of task on a runqueue to search when force migrating between
domains
GTS Tuning
26. • Linaro and ARM have been using the “interactive”
governor in their testing of the solution.
• Any governor can be used.
• b.L CPUfreq driver makes the architecture seamless to the governor.
• Example of interactive governor tuneables:
• hispeed_freq and go_hispeed_load
• target_loads
• timer_rate and min_sample_time
• above_hispeed_delay
• Governors will have tuneable parameters.
• Regardless of the governor used, there are parameters to adjust in
order to yield the right behavior
• Default values are usually not what you want
CPUFreq Governor Tuning
27. • As Linaro assimilate MP patches in the LSK, continuous
integration testing is done daily to catch possible
regressions.
• We run bbench with an audio track in the background -
good average test case.
• exercises both big and LITTLE clusters
• All automated in our LAVA environment and results
verified each day.
• Full WA regression tests with each monthly release.
• TC2 is the only b.L platform being tested at Linaro - we’d
welcome other platforms.
MP Testing at Linaro
29. More about Linaro Connect: http://connect.linaro.org
More about Linaro: http://www.linaro.org/about/
More about Linaro engineering: http://www.linaro.org/engineering/
Linaro members: www.linaro.org/members