**** BEGIN LOGGING AT Thu Aug 03 03:00:02 2017 Aug 03 07:32:12 morning all Aug 03 11:01:45 morning Aug 03 11:03:33 while looking at the patches we have in tcpdump yesterday I found that we disable the check for libpcap to "fix" a configure error Aug 03 11:04:29 the patch is file://unnecessary-to-check-libpcap.patch \ Aug 03 11:05:30 -AC_LBL_LIBPCAP(V_PCAPDEP, V_INCLS) Aug 03 11:05:32 +#AC_LBL_LIBPCAP(V_PCAPDEP, V_INCLS) Aug 03 11:07:11 the error we have witout this patch is "ERROR: /usr/bin/pcap-config should not be used, use an alternative such as pkg-config" Aug 03 11:10:37 sounds like the macro uses pcap-config and we disable that because its broken and you should use pkgconfig Aug 03 11:11:29 yes, we tried using pkg-config but it still would require a patch to tcpdump that I don't believe would be accepted Aug 03 11:11:49 well you can't use pcap-config because its broken, read the patch Aug 03 11:12:04 the comment suggests that it doesn't matter anyway Aug 03 11:12:16 so we added a depends to pkg-config to tcpdump recipe and the this extra oeconf: Aug 03 11:12:20 +EXTRA_OECONF += "\ Aug 03 11:12:21 + ac_cv_path_PCAP_CONFIG="pkg-config libpcap" \ Aug 03 11:12:22 also a change to use pkgconfig is accepted by most reasonable upstreams Aug 03 11:12:26 erm, no Aug 03 11:12:59 delete AC_LBL_LIBPCAP blaa blaa with macros from pkg.m4 Aug 03 13:35:18 is there going to be a OEDEM at the ELCE this year? Aug 03 13:38:24 I've been told yes.. Aug 03 13:38:46 I believe they said -before-... but I'm not sure I'm remembering properly Aug 03 13:38:53 Yes Aug 03 13:38:59 Do I need to send reminder Aug 03 13:39:01 Sunday before Aug 03 13:39:05 that would be good.. ;) Aug 03 13:39:09 OK Aug 03 13:39:34 ok, then i can't participate Aug 03 13:39:44 damn Aug 03 13:39:47 sorry about that Aug 03 13:40:01 i'll only arrive on monday Aug 03 13:40:11 THe M-W conf schedule is awkward Aug 03 13:43:51 Crofton|work, the date is fixed? Aug 03 13:44:03 Yes Aug 03 13:44:27 YP dev day will be Thurs and having OEDEM on Friday would be "worse" Aug 03 13:44:45 We sis announce ages ago, at least the dates Aug 03 13:45:00 There is also overlap with some rt summit Aug 03 13:45:06 so many thins to fit in a week Aug 03 13:45:35 * fray thought RT was solved.. adjust your expectations.. use good drivers.. no RT problems.. :) Aug 03 13:45:51 lol Aug 03 13:45:56 will anyhting ever be done Aug 03 13:59:37 nope. think job security! Aug 03 17:19:40 tlwoerner: I have got chromium building with clang on x86_64 Aug 03 17:20:11 tlwoerner: sent one patch for meta-browser, feel free to test it on other arches Aug 03 17:48:05 ah rpi3 build worked too Aug 03 17:48:29 clang compiles it in ~25% less time Aug 03 17:48:32 good deal Aug 03 17:56:08 fake benchmarks Aug 03 18:19:22 Crofton|work: ?? Aug 03 18:20:39 not making a benchmark statement, just an observation Aug 03 18:40:58 :) Aug 03 19:41:03 Crofton|work: Not a synthetic benchmark. But it is data. Aug 03 19:42:03 damn it fake something is a meme Aug 03 19:43:21 for the skeptics I am posting data from buildstats for do_compile Aug 03 19:43:36 gcc - Elapsed time: 1950.25 seconds Aug 03 19:43:44 clang: Elapsed time: 1694.49 seconds Aug 03 19:48:00 khem: Data! Aug 03 19:48:03 thats ~15% in do_compile Aug 03 19:48:30 To make it a stat you'd have to repeat both a few times Aug 03 19:49:06 And state your CPU/memory/disk technology and speed, etc Aug 03 19:49:20 its same machine Aug 03 19:49:24 Sure. Aug 03 19:49:26 doing nothing else Aug 03 19:49:35 But on a different machine you might get different results Aug 03 19:49:37 and I have approx 35 builds Aug 03 19:50:01 so I dont have to run many more Aug 03 19:50:57 khem: I'm comfortable with saying it's faster, but by how much depends on a lot of variables. It won't be exactly 15% fast on all build machines is my point Aug 03 19:52:11 behanw: sure, i thin the gap will just widen as the resources get tighter Aug 03 19:52:38 this machine is 64G RAM/SSD/44-cores Aug 03 19:52:47 khem: Nice.... Aug 03 19:54:00 yeah range would vary I have seen some packages show 30% Aug 03 19:56:42 what are you testing to get that stat? Aug 03 20:01:36 chromium:do_compile Aug 03 20:05:24 and I know clang is not built with best opts for speed, there are few things to tighten up Aug 03 20:06:02 we could get even better numbers Aug 03 20:07:48 e.g. DLLVM_ENABLE_ASSERTIONS=OFF -DLLVM_ENABLE_EXPENSIVE_CHECKS=OFF would match our gcc settings Aug 03 20:09:09 ahh Aug 03 20:09:28 if you can get me steps, I can check how that works against 144 threads, spinning rust and 256 GB ram Aug 03 20:11:19 man, I really derailed things :) Aug 03 20:24:58 perfect.. I found it Aug 03 20:25:11 (wrong channel) Aug 03 20:28:44 fray: add meta-clang to your layermix and add TOOLCHAIN ?= "clang" in chromium recipe or local.conf Aug 03 20:28:55 if you want clang globally Aug 03 20:29:11 so oe-core + clang + ? Aug 03 20:29:17 'er.. I mean meta-clang Aug 03 20:29:18 yes Aug 03 20:29:26 all it needs is oe-core Aug 03 20:29:27 chomium is where? Aug 03 20:29:33 meta-browser Aug 03 20:29:50 master presumably? Aug 03 20:29:56 yep all master Aug 03 20:30:56 * fray uses the new and improved bitbake-layers layerindex-fetch .... Aug 03 20:32:46 ERROR: No recipes available for: Aug 03 20:32:46 /home/mhatle/git/oss/openembedded-core/meta-browser/recipes-gnome/gnome-settings-daemon/gnome-settings-daemon_2.32.1.bbappend Aug 03 20:32:52 I assume it needs meta-gnome? Aug 03 20:33:37 which distro? the default OE doesn't have opengl -- do chromium complains it can't build Aug 03 20:34:49 lol.. I forgot.. doing 144 fetches at the same time kills my connection.. Aug 03 20:34:57 apparently my ISP doesn't like that many downloads at once.. :P Aug 03 20:35:53 fray: yeah meta-browser needs layers from meta-openembedded Aug 03 20:36:07 fray: who is your ISP Aug 03 20:36:16 it brought in meta-oe on it's own.. but not meta-gnome Aug 03 20:36:37 local ISP... it's fiber to the office -- but I can kill the connection if I do too many downloads at once.. Aug 03 20:36:56 they do traffic shaping of some kind.. and if it gets overwealmed then my interactive connections get killed off.. Aug 03 20:37:24 I need to remember to lower the max threads for the fetch Aug 03 20:37:39 since upstream chromium browser has switched to using clang as default compiler there was interest in oe community to use it as well thats where I thought of enabling it Aug 03 20:38:42 my office to the ISP is at full fibre speed.. (so for my network that is theoretically 100 Mbit.. but more like 45 or so.) Aug 03 20:39:06 but I'm only paying for 14 Mbit down... so they throttle my machines on their end when I exceed the bandwidth I pay for Aug 03 23:09:53 khem, looks like mesa failed to compile Aug 03 23:09:56 I'll switch back to the regular compiler, except for chromium Aug 03 23:16:49 lol.. building clang and gcc at the same time.. load is only 76.. :/ Aug 03 23:16:59 ohh made it to 88 Aug 03 23:17:14 102.. Aug 03 23:17:34 114 Aug 03 23:17:48 fray: mesa patches you can cherrypick from ross/mut Aug 03 23:18:01 maxed out at 119 Aug 03 23:18:11 lol Aug 03 23:18:13 they should be landing in master shortly Aug 03 23:18:51 looks like gcc cross tools 120s to build.. clang about the same Aug 03 23:19:23 take the top 4 patches from http://git.yoctoproject.org/cgit/cgit.cgi/poky-contrib/log/?h=ross/mut Aug 03 23:19:54 you need something like webkit Aug 03 23:20:01 may be you can try webkitgtk Aug 03 23:20:20 lol.. webkit is a pig.. Aug 03 23:20:36 I've still not gotten to chromium... it's at 1167 of 1767 Aug 03 23:21:04 gcc_runtime do_compile 32s... Aug 03 23:30:52 if you have capacity then you can do bitbake chromium chromium-wayland firefox qtwebkit webkitgtk qtwenengine wpewebkit Aug 03 23:31:02 all at once Aug 03 23:31:41 so far chromium is up to 74 Aug 03 23:31:44 I am sure it will smolder your machine Aug 03 23:31:47 (about 32% done) Aug 03 23:32:14 for me linking chromium takes 8 mins Aug 03 23:32:32 up to 115 now Aug 03 23:33:01 ohh 126 Aug 03 23:33:12 (I rarely see it go about about 110) Aug 03 23:33:17 131 Aug 03 23:33:23 approaching 100%.. :) Aug 03 23:33:38 62% done compiling Aug 03 23:33:39 and mind you we limit the parallism Aug 03 23:33:42 in recipe Aug 03 23:33:56 what do you limit to? BB_NUMBER_THREADS or JOBS or? Aug 03 23:34:03 up to 138.83 Aug 03 23:34:05 at 65% Aug 03 23:34:19 hot 140.27 Aug 03 23:34:22 'er.. hit Aug 03 23:34:38 I bet the machine sounds like a jet engine about to take off.. ;) (luckily I'm here, and it's at the office) Aug 03 23:34:54 142.36 @ 72% compiled Aug 03 23:35:15 143.68.. almost there.. Aug 03 23:35:19 remove ${PARALLEL_MAKE} in do_compile() in chromium-browser.inc Aug 03 23:35:21 up to 78% now Aug 03 23:35:27 suddenly just dropped a lot.. Aug 03 23:35:29 and let ninja takeover parallelism Aug 03 23:35:32 down to 125 Aug 03 23:35:50 slowly climbing again.. Aug 03 23:35:52 yeah we cap it equal to number of cores Aug 03 23:36:05 133 @ 88% Aug 03 23:36:38 it's linking.. there is only a small number of threads.. Aug 03 23:36:45 ohh there we go, it just spawned a bunch.. at 91% Aug 03 23:36:55 load had dropped into the 50s Aug 03 23:37:19 98% Aug 03 23:37:22 link uses gold which will use threading Aug 03 23:37:24 loads at 114.59 again Aug 03 23:37:35 99% done.. load dropping Aug 03 23:37:45 yes here comes the linking Aug 03 23:38:06 top is showing one process at 100% CPU... Aug 03 23:38:17 4.9 GB resident Aug 03 23:38:22 6.4 Aug 03 23:38:26 7.1 Aug 03 23:38:27 7.6 Aug 03 23:38:31 lol Aug 03 23:38:46 so far it's been running about 1:!2 Aug 03 23:38:49 er.. 1:12 Aug 03 23:39:06 it needs approx 10G peak I believe Aug 03 23:39:12 it's at 8.2g now Aug 03 23:39:15 at 1:43 Aug 03 23:39:38 load is still dropping.. says CPU utilization is current 5% Aug 03 23:39:39 done Aug 03 23:39:46 took about 2 minutes to link Aug 03 23:39:58 yes thats gold at work Aug 03 23:40:20 these cores are not the fastest.. only 2.5 Ghz Xeon 3rd gen Aug 03 23:40:29 Broadwell I think? Aug 03 23:40:34 bfd linker takes almost 4 times Aug 03 23:40:46 ok Aug 03 23:41:03 no Haswell Aug 03 23:41:21 it's the v4 that are Broadwell Aug 03 23:41:44 do you generate buildstats ? Aug 03 23:42:28 if you do check tmp/buildstats/*/chromium-wayland-*/do_compile Aug 03 23:42:41 see how long it took for you Aug 03 23:43:05 I modified it and removed the PARALLEL MAKE as you sugegsted.. I can run another build there.. Aug 03 23:43:44 Started: 1501802944.97 Aug 03 23:43:45 chromium-54.0.2810.2-r0: do_compile Aug 03 23:43:45 Elapsed time: 518.78 seconds Aug 03 23:44:02 hows that compare? Aug 03 23:44:35 well your machine is 3x quicker than me Aug 03 23:44:59 now if you want to compile with gcc then add TOOLCHAIN = "gcc" into chromium recipe Aug 03 23:45:16 so you can toggle Aug 03 23:45:17 like this Aug 03 23:46:24 if you let ninja loose it really squeezes the system Aug 03 23:46:51 but I have always ended in OOMs Aug 03 23:47:01 since I did not have enough RAM Aug 03 23:47:03 going to try it.. Aug 03 23:47:07 keeps telling em it's done.. Aug 03 23:47:12 apparently -f isn't working right naymore.. Aug 03 23:47:21 you have to do -ccleansstate Aug 03 23:47:55 yup.. in progress Aug 03 23:48:06 I guess I need to file a bug that -f isn't working Aug 03 23:48:34 ok.. unpacking and patching.. then it should go off and take over the machine.. ;) Aug 03 23:48:52 3x.. yours is 44 thread? Aug 03 23:48:58 yes Aug 03 23:49:06 so it's ALMOST linear scaling Aug 03 23:49:14 (assuming all other things are equal) Aug 03 23:49:34 probably no, mine is self made box Aug 03 23:49:52 so things are not top of class due to pricing reasons Aug 03 23:49:55 ok.. it's compiling.. already up to 10% done Aug 03 23:50:10 load is climbing.. Aug 03 23:50:13 btw. just realised that gold is not default Aug 03 23:50:24 if you want to try gold you have to uncomment USEGOLD = "-Dlinux_use_gold_flags=1" Aug 03 23:50:25 I can run it again with gold.. Aug 03 23:50:32 where is that, in chromium? Aug 03 23:50:37 yes Aug 03 23:50:42 34% done.. 80 Aug 03 23:50:48 chromium-browser.inc Aug 03 23:50:51 40%.. 85 load Aug 03 23:50:59 seems to be climbing faster this time Aug 03 23:51:17 50% at 109 Aug 03 23:51:26 yep ninja is faster than gmake as they claim Aug 03 23:52:00 60% at 129 Aug 03 23:52:20 our chormium needs upgrade its version 54 which is a bit old codebase Aug 03 23:52:23 certainly feels like the compile is slowing down now Aug 03 23:53:03 ohh hit 141 load Aug 03 23:53:27 how much ram did it take Aug 03 23:53:30 during peak Aug 03 23:53:36 %Cpu(s): 93.4 us, 6.2 sy, 0.0 ni, 0.4 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st Aug 03 23:53:36 KiB Mem : 26334428+total, 40048132 free, 19438380 used, 20385777+buff/cache Aug 03 23:53:43 not sure.. it's slightly abotu 144 now.. Aug 03 23:53:47 so it's pretty much maxed Aug 03 23:53:54 load average: 145.00, 85.69, 53.69 Aug 03 23:53:57 73% done.. Aug 03 23:54:03 just got to 145.38 Aug 03 23:54:11 memory usage is pretty steady Aug 03 23:54:19 boom, they all just went away.. and now a new crop of compiles Aug 03 23:54:23 81% done Aug 03 23:54:50 85% @ 134 Aug 03 23:55:16 89% and only one process Aug 03 23:55:34 ok.. nere we go a bunch of new ones Aug 03 23:55:48 how much RAM do you have 256G Aug 03 23:55:51 ya Aug 03 23:56:05 256 GB ram, 256 swap on an SSD.. but so far swap hasn't been touched.. Aug 03 23:56:09 thats my SSD size :) Aug 03 23:56:15 this is a 512GB SSD Aug 03 23:56:28 256 for the host OS and 256 for swap Aug 03 23:56:37 ok.. it's doing the final link now Aug 03 23:57:04 if it's like before.. it has about 1:30 to go Aug 03 23:57:30 next time try with TOOLCHAIN = "gcc" Aug 03 23:57:30 (these CPUs have 45 MB of cache memory on them.. so that helps a lot on the performance side as well) Aug 03 23:57:42 should I enable gold yet? Aug 03 23:57:48 no Aug 03 23:57:57 lets see gcc vs clang first Aug 03 23:58:02 about 30 secs... Aug 03 23:58:30 there we go.. Aug 03 23:59:03 Event: TaskStarted Aug 03 23:59:03 Started: 1501804074.66 Aug 03 23:59:04 chromium-54.0.2810.2-r0: do_compile Aug 03 23:59:04 Elapsed time: 523.06 seconds Aug 03 23:59:32 so Ninja was 4 seconds SLOWER Aug 03 23:59:45 (which may be within the noise) Aug 04 00:00:53 build starting w/ gcc Aug 04 00:01:05 (Btw my target is the default i586 here.. nothing fancy) Aug 04 00:04:33 ok.. compile started.. gcc w/ ninja Aug 04 00:04:39 52% @ 133 load.. :) Aug 04 00:07:34 70% at 151 Aug 04 00:11:19 gcc seems to be using less memory then clang.. clang was usually aroung 40 GB free.. GCC was mostly around 56 or so.. Aug 04 00:11:32 currently running the final link... Aug 04 00:12:53 102542 mhatle 20 0 12.885g 0.012t 1492 R 100.0 5.1 1:23.08 ld Aug 04 00:12:57 link is taking more memory.. odd Aug 04 00:13:31 lol.. I've never seen 0.013t before in top Aug 04 00:14:32 ld's up to 13 g.. 0.013t now.. Aug 04 00:14:47 and almost 3 minutes of processing.. you sure it wasn't runnign gold before? Aug 04 00:15:53 took 3:30 to link vs 2.. Aug 04 00:16:25 Event: TaskStarted Aug 04 00:16:25 Started: 1501804819.55 Aug 04 00:16:25 chromium-54.0.2810.2-r0: do_compile Aug 04 00:16:25 Elapsed time: 808.88 seconds Aug 04 00:17:06 so almost 5 minutes longer.. Aug 04 00:20:03 re-running... gcc + gold Aug 04 00:24:41 fray: so you have 35% gain with clang Aug 04 00:25:04 ya.. trying with gold (on gcc) then will circle back to clang Aug 04 00:25:19 thats pretty good I think Aug 04 00:25:29 khem: +1 Aug 04 00:25:38 may be you can write a small script to do some 15 builds to get an avg Aug 04 00:25:58 khem: OE needs a build stats server. ;) Aug 04 00:26:08 meta-stats... Aug 04 00:26:10 yes Aug 04 00:26:14 Hmm. That's not a terrible idea Aug 04 00:26:17 oh-meta-meta Aug 04 00:26:22 LOL Aug 04 00:26:36 double-oh-meta? Aug 04 00:26:58 thats how you feel at 5p on a thu evening Aug 04 00:27:06 LOL Aug 04 00:27:40 khem: That's how I feel as I browse downtown SFO hotels for a last minute trip in 2 weeks... Aug 04 00:28:27 oh its nice down here now a days a bit hotter to my taste Aug 04 00:28:37 92F today Aug 04 00:28:56 khem: Gah. I can't think above 27C ;) Aug 04 00:29:09 SF is almost 5C cooler always Aug 04 00:29:11 My CPU throttles. Aug 04 00:29:12 I am in SJ Aug 04 00:29:15 I know. Aug 04 00:29:44 LF has moved all classes from SJ to SF Aug 04 00:29:52 So I'm in SJ a lot less these days... Aug 04 00:30:45 back to linking w/ gold Aug 04 00:30:47 good for you, Aug 04 00:31:06 behanw: I like SFO, once I am there, I just hate the going part Aug 04 00:31:13 khem: +1 Aug 04 00:31:13 lol Aug 04 00:31:30 link time was about 1 minute Aug 04 00:31:34 so hope Elon's hyper loop will be materialized soon enough Aug 04 00:31:42 khem: Agreed. Aug 04 00:31:49 wait no.. wasn't done Aug 04 00:31:54 fray: yep, thats in line with my expectation Aug 04 00:32:01 fray: But wait... There's more! Aug 04 00:32:11 fray: btw lld ( linker from llvm project ) is even quicker than gold Aug 04 00:32:26 khem: I still haven't played with it much Aug 04 00:32:28 19255 mhatle 20 0 15.374g 0.012t 6.856g R 84.2 5.0 1:37.58 ld.gold Aug 04 00:32:35 still taking a hell of a lot of memory for linking.. Aug 04 00:32:40 khem: How does it compare to gold? Aug 04 00:33:11 fray: I can't imagine anything is slow on your compile monster. Aug 04 00:33:57 behanw: in some cases its twice faster Aug 04 00:34:02 Event: TaskStarted Aug 04 00:34:03 Started: 1501805973.29 Aug 04 00:34:03 chromium-54.0.2810.2-r0: do_compile Aug 04 00:34:03 Elapsed time: 708.30 seconds Aug 04 00:34:13 ya, gold took 100 seocnds (nearly 2 minutes off the link) Aug 04 00:34:16 khem: We can all hate fray from here. ;) Aug 04 00:34:20 but memory usage was similar Aug 04 00:34:25 lol Aug 04 00:34:49 fray I boast about your build machine in yocto classes Aug 04 00:34:58 "this is what you need"? Aug 04 00:35:12 I use it as an example of the build machines people should aspire to Aug 04 00:35:15 Yes. Aug 04 00:35:22 "everyone is jealous" Aug 04 00:35:27 the average 'monster' machine around 'here' is a 2x24 core (2x48 thread) box.. Aug 04 00:35:34 I traded sanity for more cores.. Aug 04 00:35:44 if we could use LTO effectively, the build times will shrink even more atleast with clang Aug 04 00:35:45 the thing sounds like a hair salon w/ all of the driers going at the same time.. Aug 04 00:35:46 fray: Sounds reasonable to me Aug 04 00:35:51 it's nearly impossible to sit in that lab.. Aug 04 00:35:57 khem: Wouldn't that be awesome? Aug 04 00:36:08 fray: you should put this box in out house Aug 04 00:36:15 fray: That's why you keep it in a separate room. Aug 04 00:36:30 try a different building.. ;) Aug 04 00:36:34 LOL Aug 04 00:36:36 my office is 2 miles away Aug 04 00:36:39 Even better Aug 04 00:36:57 fray: I thought the build machine was at your house? Aug 04 00:37:09 ok.. building clang + gold now Aug 04 00:37:29 behanw it was delivered to my house.. which lead to the problem 'how do I lift it to get it in my jeep to bring it to the office' Aug 04 00:37:37 answer.. remove all of the power supplies and drives.. Aug 04 00:37:44 That works Aug 04 00:37:45 and then still get one more person to help me carry it Aug 04 00:38:20 I should take a picture of it in my office.. you would laugh.... Aug 04 00:38:32 You should! Aug 04 00:38:37 I put some foam ceiling tiles around it to help direct the airflow.. Aug 04 00:38:48 it looks like a homeless person built a hovel... Aug 04 00:42:42 behanw: http://lists.llvm.org/pipermail/llvm-dev/2016-December/107981.html Aug 04 00:42:48 read the second last para Aug 04 00:43:05 Rui is one of lld contributors Aug 04 00:44:22 khem: That's great to hear! Aug 04 00:48:19 build done.. looking Aug 04 00:48:45 Event: TaskStarted Aug 04 00:48:46 Started: 1501807003.11 Aug 04 00:48:46 chromium-54.0.2810.2-r0: do_compile Aug 04 00:48:46 Elapsed time: 458.65 seconds Aug 04 00:48:51 it only cut 50 seconds off the build Aug 04 00:49:35 so clang + gold isn't too bad.. Aug 04 00:49:46 7 1/2 minutes Aug 04 00:50:07 fray: That is stupidly fast Aug 04 00:51:37 (I don't normally have any reason to build chromium.. so I don't have a good comparison) Aug 04 02:12:29 708.30 to 458.65 is still 35% **** ENDING LOGGING AT Fri Aug 04 03:00:00 2017