| IRC logs at
Set by AlexDaniel on 12 June 2018.
nwc10 good *, #moarvm 06:23
07:50 nebuchadnezzar left 07:52 nebuchadnezzar joined, MasterDuke left 07:55 nebuchadnezzar left 07:58 nebuchadnezzar joined 08:00 vrurg left, Util left 08:05 vrurg joined, Util joined 08:41 MasterDuke joined 08:51 zakharyas joined
jnthn morning o/ 09:05
tellable6 2020-08-29T15:26:01Z #raku-dev <tbrowder> jnthn: \o/ cro non-tls reverse proxy success! check
nwc10 \o 09:06
10:18 sena_kun joined 11:26 zakharyas left 11:45 Altai-man joined 11:48 sena_kun left
Geth MoarVM/master: 66 commits pushed by (Nicholas Clark)++, (Jonathan Worthington)++
nwc10 thanks for all the review feedback 12:23
I wasn't sure - do we let this stew for a bit *before* bumping NQP's MoarVM revision?
jnthn nwc10++
Probably better to bump and get it tested more, IMO
nwc10 Ooh, I was going to say "I now need to find a Camel to stick go-faster stripes no it"
but I think if I raid the childrens' bedrooms I have all the parts needed. 12:24
jnthn lizmat: For the weekly, maybe worth a mention of the new MoarVM hash impl merged :)
lizmat aaaah... cool
so, I guess we need a bump now :-)
jnthn Victim^WVolunteer to bump successfully obtained. :D 12:25
12:26 AlexDaniel joined, AlexDaniel left, AlexDaniel joined
lizmat working on bumping 12:26
jnthn I'm disappointed that google image search for "camel with go faster stripes" doesn't actually seem to find one :P
I wonder if one day they'll be able to use ML to synthesize images that don't actually exist yet to match one's search :) 12:27
nwc10 Well, I have found Camelia 12:29
failed to find the roll of go-faster stripes
Altai-man Wow, sounds awesome. Are there any benches around for people to boast? 12:32
nwc10 Altai-man: well, I tried "startup" and "setting compilation" 12:33
they are measurable.
I don't know if we have actual hash benchmarks. They should be more measurable :-)
Altai-man Well, benchmarks are lies anyway. 12:34
A lot of things can be measured, e.g. roast. 12:35
nwc10 I have no tea. Or coffee.
Altai-man nwc10++ # great work
12:36 travis-ci joined
travis-ci MoarVM build passed. Jonathan Worthington 'Merge branch 'A-better-hash'' 12:36
12:36 travis-ci left
lizmat core module installation appears to be ~15% faster 12:36
Altai-man lizmat, are you bumping rakudo or should I? 12:37
nwc10 OK, Interesting. "Your mileage may vary"
lizmat Altai-man: am running tests atm, when ok, will bump Rakudo 12:38
Altai-man roger
jnthn For once Travis bares good news... 12:40
Or is it bears...
nwc10 Spelling is hard - let's go C coding. 12:42
gha, I still have no tea/coffee/$other
jnthn Spelling *English* is hard. Thankfully, not all languages are quite so hard. :) 12:45
Altai-man jnthn, what is easier (by a good measure), in your opinion? 12:46
jnthn On spelling? Czech and Slovak for sure. :) 12:52
Not much else about them is easier, but the spelling at least is :)
nwc10 vowels rrrrrrrrrrrrrrrrr optional? 12:53
jnthn I never claimed the words were easy to say :P 12:54
lizmat ndd
.oO( funny how that was intended as "indeed", but now reads like "undead". what does that tell about me? )
Altai-man See no difference in spectest. Though maybe IDE eats up some of threads' clock-time. 12:56
jnthn IMO Russian spelling is frustratingly close to being much easier, but thanks to needing to know where the stress goes, it's still not so easy. 12:57
Altai-man Speaking Russian without stress sounds like a bold desire to me... 12:59
dogbert17 I'd say that spectest is ~10 precent faster now 13:12
precent is actually percent :)
Altai-man eh, how 13:13
117 seconds is my best result, usually it goes around 120 or something. So it should be a bit above 100, right?
dogbert17 I got 205 on my first run after the bump and 195 on the second (TEST_JOBS=10), it used to be closer to 220 before 13:21
Stage parse : 39.767 # is also quite nice
Altai-man I'm running with TEST_JOBS=24 and see no difference. Parse is 36.091 and it was ~37.5 before, so there is a speedup indeed. 13:24
13:26 zakharyas joined
dogbert17 Parse used to be around 42-43s for me recently 13:29
13:51 MasterDuke left
timotimo how's perl6 -e '' memory usage? 14:00
dogbert17 ogbert@dogbert-VirtualBox:~/repos/rakudo$ /usr/bin/time ./perl6-m -e '' 14:04
0.08user 0.01system 0:00.08elapsed 121%CPU (0avgtext+0avgdata 88080maxresident)k
timotimo a before value, too, per chance? 14:05
dogbert17 now you're demanding too much :)
let's see 14:06
gimme a couple of minutes and I'll rebuild the previous version 14:07
14:08 MasterDuke joined
timotimo 88k is about how much i have, yeah 14:12
how do i best look at the hashtable entries and metadata so i can see how it's doing? 14:20
dogbert17 0.08user 0.01system 0:00.08elapsed 121%CPU (0avgtext+0avgdata 89092maxresident)k
timotimo i tend to have to run it like 20 times and calculate the average 14:21
MasterDuke 86k is what i usually get (after the hash merge_
timotimo since it does fluctuate quite a bit
dogbert17 This is Rakudo version 2020.08.2-32-gfc75105fb built on MoarVM version 2020.08-6-g15a76dcb3
14:25 zakharyas left
nwc10 timotimo: call MVM_str_hash_fsck from gdb 14:43
er, yes, or, not sure
15:46 sena_kun joined 15:47 Altai-man left 15:53 MasterDuke left 17:04 domidumont joined 19:12 domidumont left 19:21 nebuchadnezzar left 19:22 nebuchadnezzar joined 19:23 zakharyas joined 19:45 Altai-man joined 19:48 sena_kun left
timotimo prefix_hashes is literally prefix the output with hash signs %) 19:54
.oO( When I said to build a prefix hash, this wasn't quite what I meant... :D )
timotimo i like looking at the visualization 19:57
i don't think there's a way to see what bits from the key are put into the metadata 19:59 20:02
a random example hash, just one of the first that aren't less than 128 items that shows up in gc_mark during a force_gc
it has a few slots that are 0xa or 0xb away from their ideal bucket 20:03
nwc10 that's a bit LTA. I wonder if a lower load factor would help a lot 20:06
also (future work) - the second planned thing is to steal the optimisation frmo the C++ design, where more bits of the hash are in the byte array used for "distance from ideal bucket" 20:07
which means that the lookup code can reject many entries with just a 1 byte read and compare 20:08
currently it needs to read things.
also, I forget what the average chain lenght of uthash could end up as
or, more "typical" chain lenghts
because walking that linked list is the same sort of work as walking the array, just with more cache misses thrown in 20:09
anyway, *load factor* is possibly LTA currently.
lizmat Extra Edition of the Rakudo Weekly News:
nwc10 but only *possibly* - the various folks benchmarking hashmaps in C++ seemed to be happy with high load facters and hence implied long walks 20:10
thanks for digging into this.
I was (somewhat) "it compiles - ship it"
(this is not true. It compiles and passes all the pesky tests. Finally. 20:11
"maybe ship it?"
lizmat it has been shipped and the sirens are whailing :-)
nwc10 rah. I have found my daft $ork bug. (It did not get shipped yet because it did not pass the "golden result" test) 20:12
timotimo just put my second monitor on my desk again 20:38
now to go crawling under the table and do some cabling
20:41 zakharyas left 20:47 brrt joined
lizmat And another Rakudo Weekly News hits the Net: 20:52
timotimo my computer froze up :< 21:00
21:15 AlexDaniel left 21:16 AlexDaniel joined 21:17 AlexDaniel left, AlexDaniel joined 21:20 AlexDaniel left
jnthn lizmat++ # weekly 21:29
timotimo nwc10: oh the distance-from-optimal byte doesn't have any of the hash bits in it yet? neato. that'll be cool. 21:34
brrt lizmat++ 21:40
timotimo with round-robin hashing entries in the hash will move, does that mean the little optimizations we've recently put into spesh/jit have to be tossed? 21:45
brrt what little optimizations did you put 21:46
timotimo we did spesh-time lookups of hash entries for some things 21:52
brrt ah
timotimo under the assumption that some things don't move
brrt hmm
then we need to guard against deletions I guess
and insertions, because insertions would mess with that, too 21:53
timotimo not sure if we're actually pointing at a hash slot per se 22:02
since it would also have to have been safe against insertions already
brrt indeed
timotimo there was a lot of discussion when it was developed, and i didn't pay nearly enough attention to it 22:03
[Coke] - should this be moved to rakudo/rakudo repo? 22:05
22:12 brrt left
[Coke] anyone mind if I close for lack of response? 22:28
Geth ¦ MoarVM: coke self-assigned Error!!! 22:29
¦ MoarVM: coke self-assigned closing of old issues 41, 54, 60, 66, maybe 67 ? 22:30
[Coke] seen gerd 22:46
.seen gerd
tellable6 [Coke], I haven't seen gerd around, did you mean guer?
23:17 Altai-man left