[CakeML-Dev] Bootstrap evaluation timings
Yong Kiam
tanyongkiam at gmail.com
Thu Nov 17 22:26:27 UTC 2016
Argh, that's a shame. Would moving more of asm_ok up help? There are still
some things in LabAsms that can be removed.
I'm not sure know why word_to_stack -specifically- would slow down by so
much. Could it be a one-off thing or did you hit that slowdown multiple
times already?
Another possibility: changes to the default compiler configuration could
have slowed it down.
For example this:
https://github.com/CakeML/cakeml/commit/98188addc34e5c3a034572c10eca7b3b41aff71c
(e.g. maybe more things got inlined)
On Thu, Nov 17, 2016 at 5:09 PM, Ramana Kumar <Ramana.Kumar at cl.cam.ac.uk>
wrote:
> I will put the timings from various runs of the bootstrap on this web page
> so we can more easily refer to them: https://cakeml.org/bootstrap-
> timing.txt.
> I've added partial timings for a recent revision (still waiting on the
> rest).
> There is currently a problem with word_to_stack, which is now taking 5
> hours where it used to take 1.
> And given that it seems to be stuck on "sec_ok" I'm wondering how much the
> asm_ok Asm check removal has really helped... I guess it was supposed to
> bring it down to 6h, so we'll see...
>
> On 29 September 2016 at 10:47, Ramana Kumar <Ramana.Kumar at cl.cam.ac.uk>
> wrote:
>
>> I have tried running the "sec_ok" part of the bootstrap with the asm_ok
>> checks for Asm lines turned off. Two things to report:
>>
>> The time is significantly improved:
>> runtime: 6h02m19s, gctime: 30m06s, systime: 3m11s.
>> (compared to previously: 10h46m09s, gctime: 1h07m37s, systime:
>> 5m21s.)
>>
>> Secondly, this timing information is bogus: the wall clock time was
>> definitely under 2 hours (I didn't check precisely). I expect timing for
>> threads in parallel execution are added sequentially. So for the parts of
>> the bootstrap run in parallel, the timings probably need to be divided by
>> up to 8 to get an accurate picture. Perhaps I should annotate which ones
>> are in parallel and which not.
>>
>> I think it's worthwhile removing the asm_ok checks. I can help doing so
>> if someone else sets up some cheated theorems to fill in.
>>
>> On 8 September 2016 at 14:33, Ramana Kumar <Ramana.Kumar at cl.cam.ac.uk>
>> wrote:
>>
>>> Here are some timings now including the last step (x64Bootstrap).
>>> (The commit number doesn't exactly match the output labels, since I
>>> changed some of that (about to be committed) in the middle.)
>>>
>>> Here we can see that the biggest problem (>10 hours) is the checking
>>> that all the lines in the final ASM code satisfy line_ok_light (i.e., the
>>> "sec_ok" thing below). But there are a few steps that take 1-2 hours each,
>>> and then a few more that take 20-40 mins, so ... definitely more than an
>>> overnight job still...
>>>
>>> 1a939d3
>>>
>>> eval to_mod: runtime: 1m49s, gctime: 3.4s, systime: 0.35667s.
>>> eval to_con: runtime: 8m56s, gctime: 5.5s, systime: 0.84000s.
>>> eval to_dec: runtime: 7.4s, gctime: 2.4s, systime: 0.15667s.
>>> eval to_exh: runtime: 3m32s, gctime: 8.9s, systime: 0.68667s.
>>> eval to_pat: runtime: 14.1s, gctime: 1.4s, systime: 0.06667s.
>>> eval to_clos: runtime: 5.3s, gctime: 1.1s, systime: 0.05667s.
>>> eval to_bvl: runtime: 2m36s, gctime: 12.6s, systime: 0.88000s.
>>> eval to_bvi:
>>> ... inline: runtime: 6.1s, gctime: 1.2s, systime: 0.09333s.
>>> ... optimise: runtime: 24m53s, gctime: 1m53s, systime: 17.1s.
>>> ... compile: runtime: 2m27s, gctime: 11.9s, systime: 1.5s.
>>> eval to_data: runtime: 8m11s, gctime: 21.4s, systime: 1.2s.
>>> eval data_to_word: runtime: 8m28s, gctime: 30.4s, systime: 2.4s.
>>>
>>> word_to_word: runtime: 1h00m52s, gctime: 5m53s, systime: 34.1s.
>>>
>>> clash: runtime: 5m06s, gctime: 22.3s, systime: 2.8s.
>>> (external oracle) runtime: 2m31s, gctime: 1.4s, systime: 0.77667s.
>>>
>>> chunk: runtime: 1h08m52s, gctime: 6m06s, systime: 28.2s.
>>> (chunk results) runtime: 11m49s, gctime: 1m13s, systime: 7.2s.
>>>
>>> eval word_to_stack: runtime: 1h35m38s, gctime: 3m53s, systime:
>>> 9.4s.
>>>
>>> stack_alloc: runtime: 23m27s, gctime: 45.7s, systime: 6.7s.
>>> stack_remove: runtime: 19m15s, gctime: 21.5s, systime: 4.7s.
>>> stack_names: runtime: 15m20s, gctime: 24.9s, systime: 3.2s.
>>> stack_to_lab: runtime: 42m12s, gctime: 9m27s, systime: 10.4s.
>>>
>>> filter_skip: runtime: 3m56s, gctime: 24.7s, systime: 3.2s.
>>> ffi_limit: runtime: 2m33s, gctime: 1m39s, systime: 3.2s.
>>> enc_sec: runtime: 25m01s, gctime: 1m40s, systime: 11.0s.
>>> sec_length: runtime: 4m11s, gctime: 48.9s, systime: 3.5s.
>>>
>>> compute_labels: runtime: 1h17m12s, gctime: 13m50s, systime: 58.7s.
>>>
>>> enc_secs_again: runtime: 2h00m15s, gctime: 24m27s, systime: 1m46s.
>>>
>>> upd_lab_len: runtime: 5m51s, gctime: 55.9s, systime: 3.4s.
>>> sec_length2: runtime: 3m32s, gctime: 8.4s, systime: 1.2s.
>>>
>>> compute_labels2: runtime: 1h27m09s, gctime: 22m29s, systime:
>>> 1m17s.
>>>
>>> enc_secs_again2: runtime: 2h37m41s, gctime: 1h03m15s, systime:
>>> 3m01s.
>>>
>>> pad_sectionruntime: 8m47s, gctime: 35.8s, systime: 3.8s.
>>>
>>> sec_ok: runtime: 10h46m09s, gctime: 1h07m37s, systime: 5m21s.
>>>
>>> all_secs_ok: runtime: 4.5s, gctime: 0.07667s, systime: 0.00333s.
>>> prog_to_bytes: runtime: 7m27s, gctime: 2m23s, systime: 6.3s.
>>>
>>> flat_bytes: runtime: 45m21s, gctime: 8m25s, systime: 8.2s.
>>> expand_defs: runtime: 49m25s, gctime: 31m53s, systime: 1.6s.
>>>
>>> On 7 September 2016 at 08:43, Ramana Kumar <Ramana.Kumar at cl.cam.ac.uk>
>>> wrote:
>>>
>>>> More timing for a later commit. I believe this one includes Magnus's
>>>> changes to add compile_seqs, which seem not to have had much effect. (I'm
>>>> hoping I haven't messed up the timing somehow... it's surprising how
>>>> stable they are...)
>>>>
>>>> commit: 83b3aaf
>>>>
>>>> eval to_mod: runtime: 1m50s, gctime: 5.8s, systime: 1.0s.
>>>> eval to_con: runtime: 8m41s, gctime: 4.2s, systime: 0.65000s.
>>>> eval to_dec: runtime: 11.5s, gctime: 6.4s, systime: 0.46333s.
>>>> eval to_exh: runtime: 3m27s, gctime: 6.8s, systime: 0.89000s.
>>>> eval to_pat: runtime: 14.7s, gctime: 1.7s, systime: 0.11667s.
>>>> eval to_clos: runtime: 5.5s, gctime: 1.1s, systime: 0.06333s.
>>>> eval to_bvl: runtime: 2m29s, gctime: 5.6s, systime: 0.28333s.
>>>> eval to_bvi:
>>>> ... inline: runtime: 6.3s, gctime: 1.3s, systime: 0.06333s.
>>>> ... optimise: runtime: 23m26s, gctime: 28.8s, systime: 6.3s.
>>>> ... compile: runtime: 2m18s, gctime: 4.1s, systime: 0.25000s.
>>>> eval to_data: runtime: 8m23s, gctime: 44.3s, systime: 3.2s.
>>>> eval data_to_word: runtime: 8m23s, gctime: 39.3s, systime: 3.1s.
>>>>
>>>> word_to_word: runtime: 58m18s, gctime: 4m53s, systime: 30.1s.
>>>>
>>>> clash: runtime: 5m01s, gctime: 18.4s, systime: 2.5s.
>>>> (external oracle) runtime: 2m31s, gctime: 1.9s, systime:
>>>> 0.20667s.
>>>>
>>>> chunk: runtime: 1h01m18s, gctime: 1m13s, systime: 22.9s.
>>>> (chunk results) runtime: 14m26s, gctime: 5m57s, systime: 9.1s.
>>>>
>>>> eval word_to_stack: runtime: 1h29m05s, gctime: 4m38s, systime:
>>>> 14.4s.
>>>>
>>>> stack_alloc: runtime: 23m03s, gctime: 58.3s, systime: 5.8s.
>>>> stack_remove: runtime: 20m09s, gctime: 1m29s, systime: 6.5s.
>>>> stack_names: runtime: 15m05s, gctime: 41.6s, systime: 3.9s.
>>>> stack_to_lab: runtime: 35m11s, gctime: 2m54s, systime: 12.0s.
>>>>
>>>>
>>>> On 6 September 2016 at 10:11, Yong Kiam <tanyongkiam at gmail.com> wrote:
>>>>
>>>>> Hmm, that is surprising. I expected the main function to have a good
>>>>> number of If-s (which my change should have helped with), but it seems like
>>>>> that didn't affect the time at all.
>>>>>
>>>>> Here are two possibilities:
>>>>>
>>>>> 1) Caching lookups into the colouring function for the chunk step. The
>>>>> function is defined as an option-lookup into an oracle sptree:
>>>>> https://github.com/CakeML/cakeml/blob/9937677d34446adace9f29
>>>>> d5719131f9a5b4aeac/compiler/backend/reg_alloc/reg_allocScript.sml#L110
>>>>> I'm not sure how much that will help though, since the variable names
>>>>> should be in SSA already, so it is mostly looking up different things each
>>>>> time.
>>>>>
>>>>> 2) Make splitting more aggressive, i.e. the main function should be
>>>>> divided into more pieces so that its colouring sptree isn't too big (it
>>>>> gets larger when there are more variables too).
>>>>>
>>>>> Also, are you getting some kind of effect where every parallel thread
>>>>> is just waiting for one to finish (namely the one working on the main
>>>>> function)? 2) might help divide the work more evenly.
>>>>>
>>>>>
>>>>> On Mon, Sep 5, 2016 at 7:45 PM, Ramana Kumar <
>>>>> Ramana.Kumar at cl.cam.ac.uk> wrote:
>>>>>
>>>>>> Some more timings, for two recent commits. Not a big difference, it
>>>>>> seems.
>>>>>>
>>>>>> commit: 69ac2f9
>>>>>>
>>>>>> eval to_mod: runtime: 1m45s, gctime: 3.8s, systime: 0.36667s.
>>>>>> eval to_con: runtime: 8m36s, gctime: 5.5s, systime: 0.90667s.
>>>>>> eval to_dec: runtime: 7.3s, gctime: 2.3s, systime: 0.17000s.
>>>>>> eval to_exh: runtime: 3m27s, gctime: 9.5s, systime: 1.1s.
>>>>>> eval to_pat: runtime: 13.9s, gctime: 1.5s, systime: 0.10333s.
>>>>>> eval to_clos: runtime: 5.2s, gctime: 1.1s, systime: 0.07000s.
>>>>>> eval to_bvl: runtime: 2m28s, gctime: 5.5s, systime: 0.22667s.
>>>>>> eval to_bvi: ... inline: runtime: 6.0s, gctime: 1.3s, systime:
>>>>>> 0.09000s.
>>>>>> ... optimise:runtime: 23m27s, gctime: 6.9s, systime: 3.2s.
>>>>>> ... compile: runtime: 2m33s, gctime: 19.3s, systime: 1.2s.
>>>>>> eval to_data: runtime: 8m29s, gctime: 51.5s, systime: 3.4s.
>>>>>> eval data_to_word: runtime: 8m41s, gctime: 50.2s, systime:
>>>>>> 3.4s.
>>>>>>
>>>>>> word_to_word: runtime: 58m21s, gctime: 4m28s, systime: 28.0s.
>>>>>>
>>>>>> clash: runtime: 4m58s, gctime: 15.5s, systime: 2.7s.
>>>>>> (external oracle) runtime: 2m32s, gctime: 1.8s, systime:
>>>>>> 0.24000s.
>>>>>>
>>>>>> chunk: runtime: 1h13m05s, gctime: 6m06s, systime: 26.9s.
>>>>>> (chunk results) runtime: 13m31s, gctime: 2m47s, systime: 14.4s.
>>>>>>
>>>>>> eval word_to_stack: runtime: 1h32m52s, gctime: 6m12s, systime:
>>>>>> 12.7s.
>>>>>>
>>>>>> stack_alloc: runtime: 23m03s, gctime: 42.2s, systime: 5.8s.
>>>>>> stack_remove: runtime: 19m08s, gctime: 22.7s, systime: 4.3s.
>>>>>> stack_names: runtime: 15m45s, gctime: 1m05s, systime: 5.3s.
>>>>>> stack_to_lab: runtime: 41m59s, gctime: 9m22s, systime: 8.6s.
>>>>>>
>>>>>>
>>>>>> commit: 9937677
>>>>>>
>>>>>> eval to_mod: runtime: 1m51s, gctime: 5.9s, systime: 0.90000s.
>>>>>> eval to_con: runtime: 8m48s, gctime: 4.4s, systime: 0.71667s.
>>>>>> eval to_dec: runtime: 7.2s, gctime: 2.3s, systime: 0.15667s.
>>>>>> eval to_exh: runtime: 3m30s, gctime: 9.5s, systime: 1.2s.
>>>>>> eval to_pat: runtime: 14.3s, gctime: 1.6s, systime: 0.09000s.
>>>>>> eval to_clos: runtime: 5.2s, gctime: 0.95667s, systime:
>>>>>> 0.07333s.
>>>>>> eval to_bvl: runtime: 2m28s, gctime: 5.2s, systime: 0.32667s.
>>>>>> eval to_bvi: ... inline: runtime: 5.8s, gctime: 0.88000s,
>>>>>> systime: 0.05333s.
>>>>>> ... optimise: runtime: 23m24s, gctime: 5.8s, systime: 3.0s.
>>>>>> ... compile: runtime: 2m27s, gctime: 12.5s, systime: 0.78333s.
>>>>>> eval to_data: runtime: 8m18s, gctime: 34.2s, systime: 3.0s.
>>>>>> eval data_to_word: runtime: 8m24s, gctime: 37.4s, systime:
>>>>>> 2.8s.
>>>>>>
>>>>>> word_to_word: runtime: 57m40s, gctime: 4m56s, systime: 30.1s.
>>>>>>
>>>>>> clash: runtime: 5m12s, gctime: 33.6s, systime: 3.9s.
>>>>>> (external oracle) runtime: 2m31s, gctime: 1.8s, systime:
>>>>>> 0.15667s.
>>>>>>
>>>>>> chunk: runtime: 1h07m54s, gctime: 6m07s, systime: 29.5s.
>>>>>> (chunk results) runtime: 13m21s, gctime: 2m54s, systime: 14.0s.
>>>>>>
>>>>>> eval word_to_stack: runtime: 1h34m00s, gctime: 7m32s, systime:
>>>>>> 8.1s.
>>>>>>
>>>>>> stack_alloc: runtime: 22m35s, gctime: 18.7s, systime: 3.7s.
>>>>>> stack_remove: runtime: 19m04s, gctime: 16.7s, systime: 3.9s.
>>>>>> stack_names: runtime: 15m13s, gctime: 37.8s, systime: 4.0s.
>>>>>> stack_to_lab: runtime: 36m21s, gctime: 3m29s, systime: 13.9s.
>>>>>>
>>>>>>
>>>>>> On 1 September 2016 at 09:04, Ramana Kumar <Ramana.Kumar at cl.cam.ac.uk
>>>>>> > wrote:
>>>>>>
>>>>>>> Also, updated timings for the middle bit, as of 4d0d56b (which is
>>>>>>> also for the front half timings in my previous email).
>>>>>>>
>>>>>>> eval data_to_word: runtime: 7m42s, gctime: 35.7s, systime:
>>>>>>> 2.7s.
>>>>>>>
>>>>>>> word_to_word: runtime: 56m07s, gctime: 4m48s, systime: 28.7s.
>>>>>>>
>>>>>>> clash: runtime: 4m48s, gctime: 16.0s, systime: 2.2s.
>>>>>>>
>>>>>>> (external oracle) runtime: 2m21s, gctime: 1.7s, systime:
>>>>>>> 0.10667s.
>>>>>>>
>>>>>>> chunk: runtime: 1h10m00s, gctime: 6m06s, systime: 27.1s.
>>>>>>>
>>>>>>> (use chunk) runtime: 12m26s, gctime: 2m29s, systime: 11.0s.
>>>>>>>
>>>>>>> eval word_to_stack: runtime: 1h29m28s, gctime: 7m23s,
>>>>>>> systime: 8.0s.
>>>>>>>
>>>>>>> stack_alloc: runtime: 21m13s, gctime: 24.3s, systime: 3.9s.
>>>>>>> stack_remove: runtime: 18m12s, gctime: 46.6s, systime: 4.3s.
>>>>>>> stack_names: runtime: 14m08s, gctime: 32.1s, systime: 4.1s.
>>>>>>> stack_to_lab: runtime: 36m41s, gctime: 1m19s, systime: 10.6s.
>>>>>>>
>>>>>>> It doesn't look like 91d539b made a huge difference.
>>>>>>>
>>>>>>> On 1 September 2016 at 08:42, Ramana Kumar <
>>>>>>> Ramana.Kumar at cl.cam.ac.uk> wrote:
>>>>>>>
>>>>>>>> I have some timings for the front half of the backend now:
>>>>>>>> eval to_mod: runtime: 1m45s, gctime: 9.6s, systime: 0.40000s.
>>>>>>>> eval to_con: runtime: 7m00s, gctime: 5.4s, systime: 0.80667s.
>>>>>>>> eval to_dec: runtime: 7.0s, gctime: 2.3s, systime: 0.13333s.
>>>>>>>> eval to_exh: runtime: 3m18s, gctime: 4.5s, systime: 0.67000s.
>>>>>>>> eval to_pat: runtime: 15.9s, gctime: 3.7s, systime: 0.50667s.
>>>>>>>> eval to_clos: runtime: 5.0s, gctime: 1.1s, systime: 0.07000s.
>>>>>>>> eval to_bvl: runtime: 2m18s, gctime: 5.2s, systime: 0.36000s.
>>>>>>>> eval to_bvi:
>>>>>>>> ... inline: runtime: 5.8s, gctime: 1.1s, systime: 0.07667s.
>>>>>>>> ... optimise: runtime: 22m52s, gctime: 30.1s, systime: 6.1s.
>>>>>>>> ... compile: runtime: 2m05s, gctime: 6.9s, systime: 0.72000s.
>>>>>>>> eval to_data: runtime: 5m44s, gctime: 28.8s, systime: 2.1s.
>>>>>>>>
>>>>>>>> I'm not sure why there are suspicious timings for things like
>>>>>>>> stack_alloc. One possibility is that the "time" function adds up times
>>>>>>>> across all parallel threads. In that case, it would make sense to divide
>>>>>>>> the time by 8. I would need to investigate whether this is actually what's
>>>>>>>> going on...
>>>>>>>>
>>>>>>>> On 30 August 2016 at 19:36, Magnus Myreen <magnus.myreen at gmail.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> My guess is that evaluation times get slower once we pass dataLang
>>>>>>>>> because the programs become a lot larger once the data abstraction
>>>>>>>>> has
>>>>>>>>> been removed.
>>>>>>>>>
>>>>>>>>> There are however some suspicious looking timings here. I note that
>>>>>>>>> stack_alloc does near to nothing, but still takes 31 minutes to
>>>>>>>>> run.
>>>>>>>>> Here is stack_alloc's main compiler function:
>>>>>>>>>
>>>>>>>>> https://github.com/CakeML/cakeml/blob/master/compiler/backen
>>>>>>>>> d/stack_allocScript.sml#L166
>>>>>>>>>
>>>>>>>>> Cheers,
>>>>>>>>> Magnus
>>>>>>>>>
>>>>>>>>> On 30 August 2016 at 03:43, Ramana Kumar <
>>>>>>>>> Ramana.Kumar at cl.cam.ac.uk> wrote:
>>>>>>>>> > I have some information now on where slow parts of the bootstrap
>>>>>>>>> evaluation
>>>>>>>>> > are, which we could use to direct effort in case we want to
>>>>>>>>> speed things up.
>>>>>>>>> >
>>>>>>>>> > Currently I have timings for between dataLang and labLang. (I
>>>>>>>>> know the
>>>>>>>>> > phases above dataLang are mostly quick, and the ones after
>>>>>>>>> labLang are
>>>>>>>>> > mostly slow, but I will have to get actual numbers later)
>>>>>>>>> >
>>>>>>>>> > eval data_to_word: runtime: 7m48s, gctime: 21.9s,
>>>>>>>>> systime: 2.1s.
>>>>>>>>> >
>>>>>>>>> > word_to_word: runtime: 55m07s, gctime: 5m31s, systime:
>>>>>>>>> 30.9s.
>>>>>>>>> >
>>>>>>>>> > clash: runtime: 5m33s, gctime: 34.4s, systime: 4.4s.
>>>>>>>>> >
>>>>>>>>> > (external oracle) runtime: 2m30s, gctime: 12.1s, systime:
>>>>>>>>> 1.5s
>>>>>>>>> >
>>>>>>>>> > chunk: runtime: 1h04m14s, gctime: 1m37s, systime: 20.8s.
>>>>>>>>> > (use results of chunk) runtime: 11m21s, gctime: 2m00s,
>>>>>>>>> systime:
>>>>>>>>> > 10.1s.
>>>>>>>>> >
>>>>>>>>> > eval word_to_stack: runtime: 1h34m21s, gctime: 9m42s,
>>>>>>>>> systime: 9.4s.
>>>>>>>>> >
>>>>>>>>> > stack_alloc: runtime: 31m55s, gctime: 30.3s, systime:
>>>>>>>>> 5.4s.
>>>>>>>>> >
>>>>>>>>> > stack_remove: runtime: 25m43s, gctime: 1m32s, systime:
>>>>>>>>> 6.9s.
>>>>>>>>> >
>>>>>>>>> > stack_names: runtime: 22m03s, gctime: 3m01s, systime:
>>>>>>>>> 6.6s.
>>>>>>>>> >
>>>>>>>>> > stack_to_lab: runtime: 56m00s, gctime: 4m38s, systime:
>>>>>>>>> 23.8s.
>>>>>>>>> >
>>>>>>>>> > What exactly these different phases are doing can be found by
>>>>>>>>> looking at
>>>>>>>>> > https://github.com/CakeML/cakeml/blob/master/compiler/bootst
>>>>>>>>> rap/evaluation/to_lab_x64BootstrapScript.sml.
>>>>>>>>> > I think the one confusingly called "chunk" means checking the
>>>>>>>>> results of the
>>>>>>>>> > external reg. alloc oracle.
>>>>>>>>> >
>>>>>>>>> > _______________________________________________
>>>>>>>>> > Dev mailing list
>>>>>>>>> > Dev at cakeml.org
>>>>>>>>> > https://lists.cakeml.org/listinfo/dev
>>>>>>>>> >
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Dev mailing list
>>>>>> Dev at cakeml.org
>>>>>> https://lists.cakeml.org/listinfo/dev
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
> _______________________________________________
> Developers mailing list
> Developers at cakeml.org
> https://lists.cakeml.org/listinfo/developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cakeml.org/pipermail/developers/attachments/20161117/9260b5a8/attachment-0001.html>
More information about the Developers
mailing list