Use enif_consume_timeslice and don't monopolize scheduler thread #49

urbanserj · 2013-08-24T13:10:00Z

No description provided.

davisp · 2013-08-24T16:30:18Z

Wow. That's much smaller than I was expecting to convert to using the new timeslice function. I'd like to see a couple changes before merging this though.

First, don't make this a compile time switch. Just convert the whole thing over and use a single #ifdef to change the definition of consume_timeslice that return's 0 if the function doesn't exist. I'd probably put this into util.c as jiffy_consume_timeslice so that if we have to call platform dependent functions then its easier to keep the ifdef'ed functions in a single spot.

Secondly, the calculation used to give a percentage to enif_consume_timeslice is confusing to read. As I read it, its basically calculating 1 byte of encoded or decoded binary as 0.1 percent of the total timeslice. There's also no range check to make sure it ends up between 1 and 100. In theory a long string could put this way beyond 100. In theory at least the time per byte can vary significantly. I'd probably vote in favor of just estimating based on elapsed wall-clock time assuming we can find a sub-millisecond resolution clock on each platform. Or perhaps we could fallback to some bytes consumed method if not.

Third, the processed byte calculation looks subtly broken when encoding bignums (via enc_unknown) because of how the iolist resets d->i to zero. I didn't read too hard there but I didn't see any accounting for that.

All in all, this looks fairly solid. Definitely a lot cleaner than I was expecting.

urbanserj · 2013-08-25T13:11:40Z

Hi, Paul.

First, don't make this a compile time switch. Just convert the whole thing over and use a single #ifdef to change the definition of consume_timeslice that return's 0 if the function doesn't exist. I'd probably put this into util.c as jiffy_consume_timeslice so that if we have to call platform dependent functions then its easier to keep the ifdef'ed functions in a single spot.

I have removed compile time switch #ifdef TIMESLICE and replaced macro consume_timeslice by jiffy_consume_timeslice function. I have also moved resources allocations in yield.

There's also no range check to make sure it ends up between 1 and 100.

This check is not necessary because it is present at enif_consume_timeslice: https://github.com/erlang/otp/blob/OTP_R16B01/erts/emulator/beam/erl_nif.c#L1451-L1452. But i decided to add this check for not to have problems with debug builds of beam.

I'd probably vote in favor of just estimating based on elapsed wall-clock time assuming we can find a sub-millisecond resolution clock on each platform. Or perhaps we could fallback to some bytes consumed method if not.

I used getticks from cycle.h, which is MIT licensed. For platforms x86 and x86_64 it uses the instruction RDTSC. I believe that it's a very good metric in our case.

I can squash commits into one if you wish.

Do you agree with everything else?

Thanks.

davisp · 2013-08-26T05:30:40Z

@urbanserj First off, this is quite awesome. You made all the changes I requested just fine and I'm planning on merging this but I'm still trying to reason my way through the calculation for the time slice call.

I'm new to cycle.h but as I read the file itself it seems quite adamant that we shouldn't be trying to convert it to a time unit:

https://github.com/urbanserj/jiffy/blob/364b9ef7aae0ced615591b0055ba8e1ce65ae7cb/c_src/cycle.h#L40-L42

I did some googling on various time functions to see if we couldn't cover most platforms and I was reminded on Windows' terribleness with time.

I'm thinking about switching back to something along the lines of your original patch but phrasing it slightly different. For both decoding and encoding we'll add an option that is the amount of data decoded or encoded that's handled before yielding back to Erlang. Then before yielding we just call enif_consume_timeslice(env, 100) to call it a full time slice. This seems to make a lot more sense to me rather than attempting to try and play games with Erlang's idea of a time slice as a unit of time.

How does that sound?

urbanserj · 2013-08-26T21:41:42Z

1 millisecond from enif_consume_timeslice documentation is a time estimate, but not a strict claim. Number of reductions in beam is not tunable, and it's the same for various platforms, including xeon or raspberry pi, isn't it?

Indeed, cycle.h documentation says that we shouldn't treat ticks as time, but degree of number of ticks is an approximation of nanoseconds. As I sad before, it's a good metric in our case.

Then before yielding we just call enif_consume_timeslice(env, 100) to call it a full time slice.

Before first call of jiffy, process can produce 1999 reductions. If jiffy calls enif_consume_timeslice only once w/ 100, it will be bad for system responsiveness. Maybe it's a good idea to call enif_consume_timeslice on every 10 percents on first call, and than use 100 as the second argument.

For both decoding and encoding we'll add an option that is the amount of data decoded or encoded that's handled before yielding back to Erlang.

I suppose that since noone can change number of reductions in beam, this setting won't be popular.

I did some research using erlang:system_monitor(Pid, [{long_schedule, 1}]), but I haven't got any significant results yet.

davisp · 2013-08-27T02:45:36Z

On Mon, Aug 26, 2013 at 4:41 PM, Sergey Urbanovich <notifications@github.com

wrote:

1 millisecond from enif_consume_timeslice documentation is a time
estimate, but not a strict claim. Number of reductions in beam is not
tunable, and it's the same for various platforms, including xeon or
raspberry pi, isn't it?

Just being specific in that I agree 100% with you on this. My bigger point
was that because we don't have a direct equivalency that we may want some
room to adjust things in real code.

Indeed, cycle.h documentation says that we shouldn't treat ticks as time,
but degree of number of ticks is an approximation of nanoseconds. As I sad
before, it's a good metric in our case.

I didn't see anything nearly suggesting that a tick is near a nanosecond.
If we make assumptions about clock speed then as you point out we'll be off
by orders of magnitude depending on the platform (Rasberry PI vs. Desktops).

Then before yielding we just call enif_consume_timeslice(env, 100) to
call it a full time slice.

Before first call of jiffy, process can produce 1999 reductions. If jiffy
calls enif_consume_timeslice only once w/ 100, it will be bad for system
responsiveness. Maybe it's a good idea to call enif_consume_timeslice on
every 10 percents on first call, and than use 100 as the second argument.

For both decoding and encoding we'll add an option that is the amount of
data decoded or encoded that's handled before yielding back to Erlang.

This is a really good point about how much of the timeslice a process may
have used before calling Jiffy. Definitely in agreement we should make it
incremental given this.

I suppose that since noone can change number of reductions in beam, this
setting won't be popular.

Its not about changing reduction limits so much as a way for us to figure
out a way to estimate how much work is equivalent. The reduction limit is
the dependent variable and that begs for an independent variable we can use
to test things. I fully expect that very few users would make use of this
option but when people are actively investigating scheduler issues it seems
like it could give a very important tunable insight into Jiffy.

I did some research using erlang:system_monitor(Pid, [{long_schedule, 1}]),
but I haven't got any significant results yet.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/49#issuecomment-23297388
.

I personally don't deploy code to Erlang VMs where this is an issue yet so
I haven't had motivation to combat the newer Erlang scheduling dumbness. I
do plan on merging this in near future though. My motivation for having a
free variable for tweaking the reduction equivalency is so that I can merge
without caring that users can't adjust things once a version is out that
includes it.

If I'm taking too much of your time I can try and address it in the next
few days. I realize I'm being a bit painful here but its just due to me not
having a situation where I'm able to test these things directly so I'm
trying to predict ways to not upset existing users.

Thanks for your help so far, though! I'd put this off for a long time
thinking it'd be a lot bigger of a change just to satisfy the new silly
scheduler bugs.

urbanserj · 2013-08-28T19:36:55Z

Please check out some commits I've made yesterday.

davisp · 2013-08-28T19:57:57Z

@urbanserj Awesome work! If you squash this into a single commit I'll pull it down and start running it through tests on R14 and R15 and hopefully figure out a reproducible test case that demonstrates it doesn't screw up the scheduling algorithm.

urbanserj · 2013-08-28T20:33:05Z

I've squashed it into a single commit.

knutin · 2013-10-08T17:15:06Z

Hey,

I've been playing around a bit with these changes and ran into a problem causing the beam to segfault. When encoding a bignum, jiffy returns an iolist and the continuation seems to not handle this very well.

Here's how to reproduce it (at least on R16B02):

1> jiffy:encode([trunc(math:pow(2, 64)) || _ <- lists:seq(1, 1000)]).
[1]    11130 segmentation fault (core dumped)  erl -pa ebin

As a bonus, here is something very strange:

1> jiffy:encode([trunc(math:pow(2, 64)) || _ <- lists:seq(1, 960)]), ok.
ok
2> jiffy:encode([trunc(math:pow(2, 64)) || _ <- lists:seq(1, 960)]), ok.
[1]    11372 segmentation fault (core dumped)  erl -pa ebin

davisp · 2013-10-08T19:08:03Z

@knutin I can duplicate locally on R16B01 as well. Its also not deterministic and appears to be happening back in the VM after the NIF returns which usually means its a fun misuse of the APIs. I'll look at it for a bit to see if I can find it.

davisp · 2013-10-08T19:12:09Z

Interesting note, this doesn't seem to reproduce it. Granted this should hit the new resumable logic.

[jiffy:encode(trunc(math:pow(2, 64))) || _ <- lists:seq(1, 1000)], ok.

davisp · 2013-10-08T19:49:44Z

I'm seeing a couple things that look suspect as well as a couple ways to make this a bit less complex to reason through. I'll try and find some time in the next few days to get this finished up. Apologies to @urbanserj for not getting to this sooner.

urbanserj · 2013-10-08T20:24:55Z

Thank you for reporting a bug, @knutin!

Segfaults were caused by not returning e->iolist in enc_yield. I've added a patch that fixed it. In order to avoid this type of errors we must return each explotable value of ERL_NIF_TERM type on continuation.

davisp · 2013-10-08T20:39:12Z

Ah, good find!

lpgauth · 2013-11-08T21:53:47Z

Will this make into master? 0.8.6 :)

devinus · 2014-03-13T04:17:32Z

I see this as a huge problem in jiffy right now. The database driver we're using uses jiffy to decode and encode documents, and anything slightly larger is going to take longer than 1 millisecond.

davisp · 2014-03-13T05:16:02Z

@devinus Yeah, I need to get on this. Apologies to all for letting it slip as long as I have. I'm on a business trip right now but I'll try and find time to focus on it when I'm trying to kill time.

This is due to Russell Branca's notes, attached below. There has been some discussion on what versions of Erlang CouchDB should support, and what versions of Erlang are detrimental to use. Sadly there were some pretty substantial problems in the R15 line and even parts of R16 that are landmines for CouchDB. This post will describe the current state of things and make some potential recommendations on approach. It was discovered by Basho that R15* and R16B are susceptible to scheduler collapse. There's quite a bit of discussion and information in several threads [1] [2] [3] [4] [5]. So what is scheduler collapse? Erlang schedulers can be put to sleep when there is not sufficient work to occupy all schedulers, which saves on CPU and power consumption. When the schedulers that are still running go through enough reductions to pass the work balancing threshold, they can trigger a rebalance of work that will wake up sleeping schedulers. The other mechanism for sharing scheduler load is work stealing. A scheduler that does not have any work to do can steal work from other schedulers. However a scheduler that has gone to sleep cannot steal work, it has to be woken up separately. Now the real problem of scheduler collapse occurs when you take sleeping schedulers and long running NIFs and BIFs that do not report an appropriate amount of reductions. When you have NIFs and BIFs that don't report an appropriate amount of reductions, you can get into a situation where a long running function call will only show up as taking one reduction, and never hit the work balance threshold, causing that scheduler to be blocked during the operation and no additional schedulers getting woken up. I keep mentioning "NIFs and BIFs" because it's important to note that it is _not_ just user defined NIFs that are problematic, but also a number of Erlang BIFs that don't properly report reductions. Particularly relevant to CouchDB are the BIFs `term_to_binary` and `binary_to_term` which do _not_ behave properly, and each report a single reduction count, regardless of the size of the value passed to them. Given that every write CouchDB makes goes through `term_to_binary`, this is definitely not good. This problem is systemic to all versions of R15 and R16B. In R16B01, two changes were made to alleviate the problem. First, in `OTP-11163` `term_to_binary` now uses an appropriate amount of reductions and will yield back to the scheduler. The second important change was the introduction of the `+sfwi` (Scheduler Forced Wakeup Interval) flag [6] which allows you to specify a time interval for a new watchdog process to check scheduler run queues and wake up sleeping schedulers if need be. These two changes help significantly, although from what I understand, they do not fully eliminate scheduler collapse. *NOTE*: the `+sfwi` is _not_ enabled by default, you must specify a greater than zero time interval to enable this. *WE NEED TO ENABLE THIS SETTING.* We should figure out a way to conditionally add this to vm.args or some such. On a side note, Basho runs R15B01 because they backported the `+sfwi` feature to R15B01 [7] [8]. They recommend running with `+sfwi 500` for a 500ms interval. It might be worth testing out different values, but 500 seems like a good starting point. For Riak 2.0, they will be building against R16B03-1 and 17.0 as their set of patches to R16B02 landed in R16B03-1 [9] [10]. So R16B01 sorted out the scheduler collapse issues, but unfortunately it also broke monitors, which immediately disqualifies this release as something we should recommend to users. The issues was fixed in `OTP-11225` in R16B02. I don't know of any catastrophic problems on the order of those described above in either of these releases. Basho fixed a number of unrelated bugs in R16B02 [9] [10] that have since landed in R16B03-1, which indicates we should probably prefer R16B03-1 over R16B02. R16B03 is also disqualified because it broke SSL and `erl_syntax`, resulting in the patched R16B03-1. R14B01, R14B03, and R14B04 are known good stable releases of Erlang, and in my opinion the only known stable releases > R13 that don't present issues for CouchDB (I think R16B02/R16B03-1 are too new to declare stable yet). As for R14B02, there are some bad `ets` issues with that release. It's worth pointing out that there are two known bugs in R14B01, as Robert Newson explains: ``` There are two bugs in R14B01 that we do encounter, however. 1) Another 32/64 bit oops causes the vm to attempt to allocate huge amounts of ram (terabytes, or more) if it ever tries to allocate more than 2gib of ram at once. When this happens, the vm dies and is restarted. It’s annoying, but infrequent. 2) Sometimes when closing a file, the underlying file descriptor is *not* closed, though the erlang process exits. This is rare but still quite annoying. ``` The 17.0 release brings in a number of interesting changes to help the scheduler collapse situation. `OTP-11648` improves reduction cost and yielding of `term_to_binary`. It also utilizes `OTP-11388` which allows for NIFs and BIFs to have more control over when and how they are garbage collected (we should do some investigation on the usefulness of this for NIFs like Jiffy). The 17.0 release also updates `binary_to_term` in `OTP-11535` to behave properly with reductions and yielding similar to `term_to_binary`. This marks the 17.0 release as an important one for CouchDB as now `term_to_binary` and `binary_to_term` both behave properly. One other interesting item introduced in the 17.0 release is the concept of dirty schedulers [12] [13]. This is an experimental feature providing CPU and I/O schedulers specifically for NIFs that are known to take longer that 1ms to run. In general, we want to make sure the NIFs we use will yield and report reductions properly, but for situations where that isn't feasible, we may want to look into using dirty schedulers down the road when it's a non experimental feature. In my opinion we need to take the Erlang release issues more seriously than we currently do and provide strong recommendations to users on what versions of Erlang we support. I suggest we loosely take an approach similar to Debian, and make three recommendations: * OldStable: [R14B01, R14B03, R14B04 (NOTE: _not_ R14B02)] * Unstable: [R16B03-1 recommended, R16B02 acceptable] * Experimental: [17.0] I'm not suggesting permanently having three Erlang releases recommended like this, but it currently seems appropriate. I think long term we should target 17.x as our preferred Erlang release, and then make a CouchDB 3.0 release that is backwards incompatible with anything less than 17.0 so that we can switch over to using maps. The narrowness of the acceptable releases list is going to cause some problems. Debian Wheezy runs R15B01, which as established above, is not good to run with unless you have the `+sfwi` patch, and I'm sure there are many other distros running R15 and R16B or R16B01. I think it would be useful to users to have a set of packages with a proper Erlang CouchDB release allowing us to bless specific versions of Erlang and bundle it together, but I know this idea goes against the recent change in stance on working with distributions, and I don't know the ASF stance on this issue well enough to comment on the legality of it. That said, it does seem like the logical approach until we get a range of stable releases spread out through the distros. We need to make sure that all NIFs we use that could potentially take longer than 1ms to run properly yield and report reductions. For Jiffy, there is already a good start on this work [11]. We'll want to look into what needs to be done for the rest of the NIFs. There's quite a bit of information here, and plenty more in the footnotes, so I hope this gives a good overview of the current state of Erlang releases and helps us to make informed decisions on what approach to take with Erlang releases. [1] http://comments.gmane.org/gmane.comp.lang.erlang.bugs/3564 [2] http://erlang.org/pipermail/erlang-questions/2013-April/073490.html [3] http://erlang.org/pipermail/erlang-questions/2012-October/069503.html [4] http://erlang.org/pipermail/erlang-questions/2012-October/069585.html [5] http://permalink.gmane.org/gmane.comp.lang.erlang.bugs/3573 [6] http://erlang.org/pipermail/erlang-patches/2013-June/004109.html [7] https://gist.github.com/evanmcc/a599f4c6374338ed672e [8] http://data.story.lu/2013/06/23/riak-1-3-2-released [9] basho/otp@erlang:maint...OTP_R16B02_basho4 [10] https://groups.google.com/forum/#!topic/nosql-databases/XpFKVeUBdn0 [11] davisp/jiffy#49 [12] erlang/otp@c1c03ae [13] http://www.erlang.org/doc/man/erl_nif.html#dirty_nifs

davisp · 2014-06-13T22:21:54Z

Just released 0.10.1 which includes this functionality. Huge thanks to @urbanserj for the work putting together this patch. I ended up moving things around slightly but its all directly based on his work.

Use enif_consume_timeslice and don't monopolize scheduler thread

f63c10a

Fix segfault on encoding a bignum

4d57502

davisp closed this Jun 13, 2014

Use enif_consume_timeslice and don't monopolize scheduler thread #49

Use enif_consume_timeslice and don't monopolize scheduler thread #49

urbanserj commented Aug 24, 2013

davisp commented Aug 24, 2013

urbanserj commented Aug 25, 2013

davisp commented Aug 26, 2013

urbanserj commented Aug 26, 2013

davisp commented Aug 27, 2013

urbanserj commented Aug 28, 2013

davisp commented Aug 28, 2013

urbanserj commented Aug 28, 2013

knutin commented Oct 8, 2013

davisp commented Oct 8, 2013

davisp commented Oct 8, 2013

davisp commented Oct 8, 2013

urbanserj commented Oct 8, 2013

davisp commented Oct 8, 2013

lpgauth commented Nov 8, 2013

devinus commented Mar 13, 2014

davisp commented Mar 13, 2014

davisp commented Jun 13, 2014

Use enif_consume_timeslice and don't monopolize scheduler thread #49

Use enif_consume_timeslice and don't monopolize scheduler thread #49

Conversation

urbanserj commented Aug 24, 2013

davisp commented Aug 24, 2013

urbanserj commented Aug 25, 2013

davisp commented Aug 26, 2013

urbanserj commented Aug 26, 2013

davisp commented Aug 27, 2013

urbanserj commented Aug 28, 2013

davisp commented Aug 28, 2013

urbanserj commented Aug 28, 2013

knutin commented Oct 8, 2013

davisp commented Oct 8, 2013

davisp commented Oct 8, 2013

davisp commented Oct 8, 2013

urbanserj commented Oct 8, 2013

davisp commented Oct 8, 2013

lpgauth commented Nov 8, 2013

devinus commented Mar 13, 2014

davisp commented Mar 13, 2014

davisp commented Jun 13, 2014