On 02/18/2015 03:07 AM, Ingo Molnar wrote:
* Jason Baron [off-list ref] wrote:
quoted
Epoll file descriptors that are added to a shared wakeup
source are always added in a non-exclusive manner. That
means that when we have multiple epoll fds attached to a
shared wakeup source they are all woken up. This can lead
to excessive cpu usage and uneven load distribution.
This patch introduces two new 'events' flags that are
intended to be used with EPOLL_CTL_ADD operations.
EPOLLEXCLUSIVE, adds the epoll fd to the event source in
an exclusive manner such that the minimum number of
threads are woken. EPOLLROUNDROBIN, which depends on
EPOLLEXCLUSIVE also being set, can also be added to the
'events' flag, such that we round robin through the set
of waiting threads.
An implementation note is that in the epoll wakeup
routine, 'ep_poll_callback()', if EPOLLROUNDROBIN is set,
we return 1, for a successful wakeup, only when there are
current waiters. The idea is to use this additional
heuristic in order minimize wakeup latencies.
Signed-off-by: Jason Baron <redacted>
---
fs/eventpoll.c | 25 ++++++++++++++++++++-----
include/uapi/linux/eventpoll.h | 6 ++++++
2 files changed, 26 insertions(+), 5 deletions(-)
diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index d77f944..382c832 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -92,7 +92,8 @@
*/
/* Epoll private bits inside the event mask */
-#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
+#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | \
+ EPOLLEXCLUSIVE | EPOLLROUNDROBIN)
107 more quoted lines
yes, in fact we are currently working around these imbalances
by doing register/unregister (EPOLL_CTL_ADD/EPOLL_CTL_DEL),
periodically to re-set the order of the queues. This resolves the
balancing to an extent, but not the spurious wakeups.
The solution is to add this new ABI to allow epoll events
to be actively load-balanced both between the persistent
'task lists', and to also allow the individual task lists
to act as dynamic runqueues: the head of the list is likely
to be sleeping, newly woken tasks get moved to the tail of
the list.
This has two main advantages: firstly it solves the O(N)
(micro-)problem, but it also more evenly distributes events
both between task-lists and within epoll groups as tasks as
well.
Its solving 2 issues - spurious wakeups, and more
even loading of threads. The event distribution is more
even between 'epoll groups' with this patch, however,
if multiple threads are blocking on a single 'epoll group',
this patch does not affect the the event distribution there.
Currently, threads are added to 'epoll group' as exclusive
already, so when you have multiple threads blocking on
an epoll group, only one wakes up. In our use case, we
have a 1-to-1 mapping b/w threads and epoll groups, so
we don't have spurious or un-balanced wakeups there.
That suggests the alternative user-space model for
addressing this problem. That is, to have a single epoll
group added with EPOLLONESHOT. In this way threads
can pull work or events off of a single queue, work on
the event and then re-arm (such that other threads don't
see events from that source in the meantime).
This, however, means all threads work on all events, and
they all have to synchronize to an extent on the single queue.
That is all register/unregister and re-arm event to that queue
have to be visible or synchronized for all the waiters. This
model also doesn't allow userspace to partition events that
are naturally local to thread, since there a single epoll group.
The second userspace model is to have worker threads with
their own separate epoll groups, and then have separate
thread(s) to address the shared wakeup sources. Then the
threads that are waiting on the shared wakeup sources can
distribute the events fairly to the worker threads. This involves
extra context switching for shared events, and I think ends up
degenerating back into the original problem.
The disadvantages: slightly higher management micro-costs,
plus a global waitqueue list, which used to be read-mostly,
is now actively dirtied by every event, adding more global
serialization. The latter is somewhat muted by the fact
that the waitqueue lock itself is already a global
serialization point today and got dirtied by every event,
and the list head is next to it, usually in the same
cacheline.
Yes, I'm a bit concerned about the changes to the core
wakeup function, however in the non-rotate case, the
only additional write is to initialize the 'rotate_list' on entry.
I measured the latency of the __wake_up_common() for
the case where this code was added and we were not
doing 'rotate', and I didn't measure any additional latency
with ftrace, but it perhaps warrants more careful testing.
The outstanding issues I have are:
1) Does epoll need 2 new flags here - EPOLLEXCLUSIVE and
EPOLLROUNDROBIN. IE should they just be combined since
EPOLLROUNDROBIN depends on EPOLLEXCLUSIVE, or is there
a valid use for just EPOLLEXCLUSIVE (wake up the first waiter
but don't do the balancing)?
2) The concern Andy raised regarding potential starvation.
That is could a adversarial thread cause us to miss wakeups
if it can add itself exclusively to the shared wakeup source.
Currently, the adversarial thread would need to simply be
able to open the file in question. For things like pipe, this is
not an issue, but perhaps it is for files in the global
namespace...
Thanks,
-Jason