|
An unbound workqueue can be served by multiple worker_pools to improve
locality. The segmentation is achieved by grouping CPUs into pods. By
default, the cache boundaries according to cpus_share_cache() define the
CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the
system has two L3 caches. The workqueue would be mapped to two worker_pools
each serving one L3 cache domains.
While this improves locality, because the pod boundaries are strict, it
limits the total bandwidth a given issuer can consume. For example, let's
say there is a thread pinned to a CPU issuing enough work items to saturate
the whole machine. With the machine segmented into two pods, no matter how
many work items it issues, it can only use half of the CPUs on the system.
While this limitation has existed for a very long time, it wasn't very
pronounced because the affinity grouping used to be always by NUMA nodes.
With cache boundaries as the default and support for even finer grained
scopes (smt and cpu), it is now an a lot more pressing problem.
This patch implements non-strict affinity scope where the pod boundaries
aren't enforced strictly. Going back to the previous example, the workqueue
would still be mapped to two worker_pools; however, the affinity enforcement
would be soft. The workers in both pools would have their cpus_allowed set
to the whole machine thus allowing the scheduler to migrate them anywhere on
the machine. However, whenever an idle worker is woken up, the workqueue
code asks the scheduler to bring back the task within the pod if the worker
is outside. ie. work items start executing within its affinity scope but can
be migrated outside as the scheduler sees fit. This removes the hard cap on
utilization while maintaining the benefits of affinity scopes.
After the earlier ->__pod_cpumask changes, the implementation is pretty
simple. When non-strict which is the new default:
* pool_allowed_cpus() returns @pool->attrs->cpumask instead of
->__pod_cpumask so that the workers are allowed to run on any CPU that
the associated workqueues allow.
* If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets
the field to a CPU within the pod.
This would be the first use of task_struct->wake_cpu outside scheduler
proper, so it isn't clear whether this would be acceptable. However, other
methods of migrating tasks are significantly more expensive and are likely
prohibitively so if we want to do this on every work item. This needs
discussion with scheduler folks.
There is also a race window where setting ->wake_cpu wouldn't be effective
as the target task is still on CPU. However, the window is pretty small and
this being a best-effort optimization, it doesn't seem to warrant more
complexity at the moment.
While the non-strict cache affinity scopes seem to be the best option, the
performance picture interacts with the affinity scope and is a bit
complicated to fully discuss in this patch, so the behavior is made easily
selectable through wqattrs and sysfs and the next patch will add
documentation to discuss performance implications.
v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Add three more affinity scopes - WQ_AFFN_CPU, SMT and CACHE - and make CACHE
the default. The code changes to actually add the additional scopes are
trivial.
Also add module parameter "workqueue.default_affinity_scope" to override the
default scope and "affinity_scope" sysfs file to configure it per workqueue.
wq_dump.py and documentations are updated accordingly.
This enables significant flexibility in configuring how unbound workqueues
behave. If affinity scope is set to "cpu", it'll behave close to a per-cpu
workqueue. On the other hand, "system" removes all locality boundaries.
Many modern machines have multiple L3 caches often while being mostly
uniform in terms of memory access. Thus, workqueue's previous behavior of
spreading work items in each NUMA node had negative performance implications
from unncessarily crossing L3 boundaries between issue and execution.
However, picking a finer grained affinity scope also has a downside in that
an issuer in one group can't utilize CPUs in other groups.
While dependent on the specifics of workload, there's usually a noticeable
penalty in crossing L3 boundaries, so let's default to CACHE. This issue
will be further addressed and documented with examples in future patches.
Signed-off-by: Tejun Heo <tj@kernel.org>
|
