* [PATCH] epollexclusive: use maxevents=1 for epoll_wait
@ 2023-03-28 12:24 14% ` Eric Wong
0 siblings, 0 replies; 1+ results
From: Eric Wong @ 2023-03-28 12:24 UTC (permalink / raw)
To: Jean Boussier; +Cc: unicorn-public
Jean Boussier <jean.boussier@shopify.com> wrote:
> Interesting patch, I wonder if you have considered RB_ALLOCV_N, e.g:
>
> https://patch-diff.githubusercontent.com/raw/Shopify/pitchfork/pull/37.patch
Were you able to benchmark an improvement using alloca?
I would think it gets lost in the noise.
I don't like alloca in general since it can make the compiler
over-allocate stack (because default cflags favors speed over
size). It also makes static analysis more difficult.
But C Ruby is already infested with alloca, so one more/less
likely isn't meaningful. But, maybe I have a better solution
than either alloca or malloc below.
Been busy dealing with other stuff, but something new just
occured to me based on something I've abused for over a decade.
Background:
Outside of unicorn (and COMPLETELY different designs), I've
abused epoll_wait with maxevents=1 for over a decade because
it gives the most fairness in multi-threaded cases. epoll is
a queue, too, kqueue just got the more appropriate name.
It never occured to me until right now now, but maxevents=1
combined w/ EPOLLEXCLUSIVE should ALSO improve fairness
(compared to maxevents >1 + EPOLLEXCLUSIVE) in a multi-process,
single-threaded environment like unicorn if configured with
multiple listeners.
Explanation in the commit message and code comments below,
hopefully it makes sense.
Anybody willing to benchmark the following on a large SMP
machine? (I hate consumerism and refuse to have powerful HW)
---------8<--------
Subject: [PATCH] epollexclusive: use maxevents=1 for epoll_wait
This allows us to avoid both malloc (slow) and alloca
(unpredictable stack usage) at the cost of needing to make more
epoll_wait syscalls in a rare case.
In unicorn (and most servers), I expect the most frequent setup
is to have one active listener serving the majority of the
connections, so the most frequent epoll_wait return value would
be 1.
Even with >1 events, any syscall overhead saved by having
epoll_wait retrieve multiple events is dwarfed by Rack app
processing overhead.
Worse yet, if a worker retrieves an event sooner than it can
process it, the kernel (regardless of EPOLLEXCLUSIVE or not) is
able to enqueue another new event to that worker. In this
example where `a' and `b' are both listeners:
U=userspace, K=kernel
K: client hits `a' and `b', enqueues them both (events #1 and #2)
U: epoll_wait(maxevents: 2) => [ a, b ]
K: enqueues another event for `b' (event #3)
U: process_client(a.accept) # this takes a long time
While process_client(a.accept) is happening, `b' can have two
clients pending on a given worker. It's actually better to
leave the first `b' event unretrieved so the second `b'
event can go to the ep->rdllist of another worker.
The kernel is only capable of enqueuing an item if it hasn't
been enqueued. Meaning, it's impossible for epoll_wait to ever
retrieve `[ b, b ]' in one call.
---
ext/unicorn_http/epollexclusive.h | 31 +++++++++++++------------------
1 file changed, 13 insertions(+), 18 deletions(-)
diff --git a/ext/unicorn_http/epollexclusive.h b/ext/unicorn_http/epollexclusive.h
index 677e1fe..8f4ea9a 100644
--- a/ext/unicorn_http/epollexclusive.h
+++ b/ext/unicorn_http/epollexclusive.h
@@ -64,18 +64,22 @@ static VALUE prep_readers(VALUE cls, VALUE readers)
#if USE_EPOLL
struct ep_wait {
- struct epoll_event *events;
+ struct epoll_event event;
rb_io_t *fptr;
- int maxevents;
int timeout_msec;
};
static void *do_wait(void *ptr) /* runs w/o GVL */
{
struct ep_wait *epw = ptr;
-
- return (void *)(long)epoll_wait(epw->fptr->fd, epw->events,
- epw->maxevents, epw->timeout_msec);
+ /*
+ * Linux delivers epoll events in the order received, and using
+ * maxevents=1 ensures we pluck one item off ep->rdllist
+ * at-a-time (c.f. fs/eventpoll.c in linux.git, it's quite
+ * easy-to-understand for anybody familiar with Ruby C).
+ */
+ return (void *)(long)epoll_wait(epw->fptr->fd, &epw->event, 1,
+ epw->timeout_msec);
}
/* :nodoc: */
@@ -84,14 +88,10 @@ static VALUE
get_readers(VALUE epio, VALUE ready, VALUE readers, VALUE timeout_msec)
{
struct ep_wait epw;
- long i, n;
- VALUE buf;
+ long n;
Check_Type(ready, T_ARRAY);
Check_Type(readers, T_ARRAY);
- epw.maxevents = RARRAY_LENINT(readers);
- buf = rb_str_buf_new(sizeof(struct epoll_event) * epw.maxevents);
- epw.events = (struct epoll_event *)RSTRING_PTR(buf);
epio = rb_io_get_io(epio);
GetOpenFile(epio, epw.fptr);
@@ -99,17 +99,12 @@ get_readers(VALUE epio, VALUE ready, VALUE readers, VALUE timeout_msec)
n = (long)rb_thread_call_without_gvl(do_wait, &epw, RUBY_UBF_IO, NULL);
if (n < 0) {
if (errno != EINTR) rb_sys_fail("epoll_wait");
- n = 0;
- }
- /* Linux delivers events in order received */
- for (i = 0; i < n; i++) {
- struct epoll_event *ev = &epw.events[i];
- VALUE obj = rb_ary_entry(readers, ev->data.u64);
+ } else if (n > 0) { /* maxevents is hardcoded to 1 */
+ VALUE obj = rb_ary_entry(readers, epw.event.data.u64);
if (RTEST(obj))
rb_ary_push(ready, obj);
- }
- rb_str_resize(buf, 0);
+ } /* n == 0 : timeout */
return Qfalse;
}
#endif /* USE_EPOLL */
^ permalink raw reply related [relevance 14%]
Results 1-1 of 1 | reverse | options above
-- pct% links below jump to the message on this page, permalinks otherwise --
2023-03-01 23:12 [PATCH] epollexclusive: avoid Ruby object allocation for buffer Eric Wong
2023-03-22 12:53 ` Jean Boussier
2023-03-28 12:24 14% ` [PATCH] epollexclusive: use maxevents=1 for epoll_wait Eric Wong
Code repositories for project(s) associated with this public inbox
https://yhbt.net/unicorn.git/
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).