IRC Log

Thursday October 27, 2011

[Time] NameMessage
[00:51] travlr Just noticed Pieter will be on FLOSS weekly Dec. 28 ...+1
[01:01] lusis indeed
[01:01] lusis looking forward to that one
[01:01] lusis stopped listening a while ago since the interviews got rather weak
[01:11] travlr lusis: yeah sometimes they can be lame. But Twit in general is pretty cool. I'm usually watching.
[01:12] lusis I guess I got tired of some of the banal questions from Merlyn
[01:12] lusis randall...whatever he goes by ;)
[01:12] lusis Maybe the target audience isn't me
[01:12] lusis heh
[01:13] travlr i wish they got more into the programming aspects.. they need a linux today as well for hackers.
[01:14] tarcieri sup there lusis
[01:14] lusis tarcieri: hola!
[01:15] lusis tarcieri: was reading the backlog. Looking forward to seeing what you come up with
[01:15] tarcieri I'm dabbling with plugging Celluloid into 0mq
[01:15] tarcieri yeah we'll see
[01:15] lusis tarcieri: wish I had more time to help =/
[01:15] tarcieri heh
[01:15] tarcieri I might have a bit of time on my hands here
[01:16] lusis good or bad "bit of time"?
[01:17] tarcieri haha
[01:17] tarcieri depends how you look at it I guess
[01:17] lusis heh
[01:36] tarcieri okay, so I'm confused
[01:36] tarcieri when I use explicit pairs
[01:36] tarcieri is that 1:1 for a particular TCP port?
[01:45] tarcieri maybe I really want push and pull
[01:46] tarcieri each node is a message sink
[01:46] tarcieri that can have N people sending it messages
[02:00] tarcieri so a pull sock can bind
[02:01] tarcieri and N clients can push
[05:56] sustrik cremes: there?
[07:43] CIA-79 libzmq: 03Martin Sustrik 07master * rb3cda2a 10/ src/kqueue.cpp : Bug in kqueue poller fixed (issue 261) ...
[09:50] CIA-79 libzmq: 03Paul Betts 07master * r1b706ac 10/ (src/err.cpp src/err.hpp): Enable exceptions raising on assert on Win32 ...
[09:51] CIA-79 libzmq: 03Martin Sustrik 07master * r68ab5f8 10/ AUTHORS : Paul Betts added to the AUTHORS file ...
[11:04] mikko sustrik: gj on the kqueue stuff
[11:04] mikko sustrik: looking at the UApycon push/pull convo
[11:04] mikko does this tie in with libzmq-160
[11:05] sustrik thanks
[11:05] sustrik yes, it's basically the same thing
[11:05] sustrik however, decent shutdown requires some API changes as well
[11:06] sustrik some kind of 2-phase shutdown
[11:06] sustrik 1. ask socket to shutdown
[11:06] sustrik 2. read the remaining queued messages
[11:06] sustrik 3. exit
[11:06] sustrik so, i guess, first we should decide on what exactly should be done and how
[11:12] mikko sustrik: this shutdown business might tie in with the dealer lost messages case
[11:13] sustrik yes, it does
[11:13] sustrik it doesn't solve the case of byzantine failure
[11:13] sustrik but no messages should be lost on decent shut down
[11:20] mikko sustrik: this is true
[11:34] jond sustrik: hi, what was the underlying issue with the kqueue? the patch seems to protect against multiple add/rm of fd?
[11:46] cremes sustrik: just got up... i'll be around in about 90 minutes if u still need me.
[11:56] sustrik jond: yes
[11:56] sustrik there was a case where a fd was removed *twice* from the pollset
[11:56] sustrik which caused the error
[11:57] sustrik cremes: everything's done already
[11:57] sustrik thanks for providing the osx box
[12:00] cremes sustrik: you are welcome; glad i could be of service
[12:00] cremes now i don't have to compile on osx with FORCE_POLL or whatever the macro is!
[12:01] sustrik great
[12:01] sustrik if there's another problem with osx in the future i'll ping you
[12:02] CIA-79 libzmq: 03Ben Gray 07master * r9e000c8 10/ (src/dist.cpp src/msg.cpp src/msg.hpp): Patch for issue LIBZMQ-275. Dealing with VSM in distribution when pipes fail to write. ...
[12:03] CIA-79 libzmq: 03Martin Sustrik 07master * r9b3e61a 10/ AUTHORS : Ben Gray added to the AUTHORS file ...
[12:06] cremes ok
[14:00] nbx hello everyone, i've a problem. i installed 0MQ 2.1.10 on debian linux distribution, however after everything's done and installed - i cannot run any of the tests bundled in /perf directory. programs just don't respond after i run them, they hang. how to verify if i did something wrong while installing 0MQ?
[14:03] sustrik nbs: what command lines are you using to run them?
[14:03] cremes nbx: there should be a 'make test' or similar target for make that will try to run the tests
[14:07] nbx sustrik i change to /perf directory and run the command from the example with ./local_lat tcp://eth0:5555 1 100000
[14:07] sustrik what about the other peer?
[14:07] nbx cremes, the installation says that performance tests executables will be available in /perf directory after buliding and they are..
[14:08] cremes nbx: you were asking how to make sure the lib built correctly; there is a make target that will run some
[14:08] cremes tests that *use* the library and confirm it was built correctly
[14:08] nbx sustrik it says this is the local latency test, i'm not sure what other peer.. i literally copy pasted the example from 0mq website
[14:08] cremes the perf tests are a separate isue
[14:08] nbx aha i see
[14:08] cremes nbx: give us the url that you copied the example from
[14:09] cremes there is usually a local and a remote (2 peers)
[14:09] nbx http://www.zeromq.org/results:perf-howto
[14:09] nbx i ran the 1st command that tests local latency
[14:09] cremes wrong
[14:09] sustrik have a look two lines below that :)
[14:09] cremes read the text there... it tells you exactly what to do
[14:11] nbx now i feel dumb :) thanks for the help
[14:12] cremes nbx: don't feel dumb!
[14:13] nbx now the second question, is there any reason i should be getting segmentation fault errors when trying to connect a client to 0mq server? again, i used PHP examples from the introduction article
[14:13] cremes we are kind of "tough" on people to read the docs thoroughly, that's all
[14:13] nbx and how to determine why i got them?
[14:13] cremes segfaults are bad
[14:13] cremes you should not get them
[14:13] nbx well, i've been into message queues for about 2 weeks now, i do read everything usually but after the whole day of reading.. you miss things :)
[14:13] mikko nbx: does php segfault?
[14:14] nbx php script reports segfault, but the 0mq client fails to connect to 0mq server
[14:14] mikko hmm, sounds odd
[14:14] mikko can you try this:
[14:14] mikko gdb --args php script.php
[14:14] mikko then type run
[14:14] mikko and when it crashes do 'bt'
[14:15] mikko and put the backtrace to gist.github.com
[14:15] mikko also, what version of PHP and which OS?
[14:15] nbx php 5.3.8 and debian 6
[14:15] mikko have you got imagick or uuid extensions installed?
[14:16] nbx i got imagick
[14:16] mikko ok, thats it
[14:16] nbx alright, killing the imagick then
[14:16] mikko this is an old bug in glibc
[14:17] mikko you can get it to work by loading imagick after zmq
[14:17] nbx alright, going to try it, thanks a lot guys :) (i'll be back probably, this 0mq really owns)
[14:17] mikko this is a bug in initialising thread-local storage which causes a crash in libuuid
[14:17] mikko np
[14:23] nbx mikko that took care of it, thanks
[15:12] CIA-79 libzmq: 03Bernd Prager 07master * r52bab42 10/ src/zmq.cpp : Missing bracket added ...
[15:32] cremes i need someone to explain to me the functions zmsg_wrap and zmsg_unwrap in the czmq library
[15:32] cremes is the intention for these calls to add and remove envelope information for XREP sockets?
[15:33] mikko pieter is your man
[15:33] mikko i would guess
[15:33] cremes he's in seoul... i hope someone else is using the czmq lib and can help me :)
[15:34] cremes i'm not confused by the code... i'm confused by the api as in "why does it exist?"
[16:09] sustrik cremes: i guess asking on list could help
[16:09] cremes writing a message now :)
[16:09] sustrik pieter may get to the list even in seoul
[16:09] sustrik :)
[16:12] mikko cremes: did you look at the source?
[16:12] cremes mikko: i did; i understand the source completely
[16:12] cremes it's not the mechanics of it that have me confused
[16:12] cremes it's the purpose of it
[17:08] cremes sustrik: in 3.x, would it make sense for the zmq_msg_t to be modified so it also carries a flag indicating label true/false?
[17:09] cremes e.g. int zmq_msg_label(zmq_msg_t *msg)
[17:09] cremes returns 0 for false, 1 for true
[17:09] cremes that would be instead of using getsockopt to test whether the last message received was a label
[17:10] cremes that makes more sense to me; thoughts?
[18:08] indygreg cremes: I dig your recent 3.x msg API change proposal. +100 from me
[18:09] cremes ha! are you gregory szorc?
[18:09] indygreg yup
[18:09] cremes i'm just reading your reply to the other thread... the idea of returning an int and testing the bits
[18:09] cremes is interesting
[18:09] cremes i hadn't considered that
[18:09] indygreg ... very similar to flagmatch
[18:10] indygreg except more C like
[18:10] cremes yes
[18:10] cremes i like the "less C like" version :)
[18:10] cremes but i could be happy with something like:
[18:10] indygreg well, that's why you have high-level languages or multiple APIs to be more explicit
[18:10] cremes int zmq_msg_flags(zmq_msg_t *msg)
[18:10] cremes and testing the result
[18:11] cremes right
[18:11] indygreg that works
[18:12] cremes looking through socket_base.cpp, i think something like zmq_msg_flags() would be a pretty easy replacement
[18:12] cremes for the current methodology
[18:14] cremes indygreg: do you like the name zmq_msg_flags() or is there something more idiomatic for C?
[18:21] indygreg cremes: zmq_msg_flags() sounds good to me! zmq_msg_fields() would be another choice. but flags is better IMO since I think "flag" in C implies bits
[18:22] cremes ok
[18:22] cremes i will reply to my own thread and modify the proposal
[18:30] indygreg I also think separate int zmq_msg_flags(zmq_msg_t *msg) and int zmq_msg_flag(zmq_msg_t *msg, int flag) would be a decent addition - at the risk of having some redundancy in the API
[18:31] indygreg cremes: ^^^
[18:32] cremes hmmm... you won me over to the "flags" api... why add this second one?
[18:32] LodeRunner hello - does creating an inproc socket allocate a file descriptor?
[18:32] indygreg LodeRunner: AFAIK inproc sockets are basically in-memory data structures
[18:32] indygreg so no
[18:33] LodeRunner indygreg: that's what I thought, but I'm getting "too many open files" and I see tons of 'unix' sockets in lsof when I raise the number of threads in my app
[18:34] indygreg cremes: I would prefer zmq_msg_flags() only, but some may not like the low-level nature of it. do you force people to do "if (zmq_msg_flags(msg) & ZMQ_MSG_MORE)"? do you write macros: #define TEST_ZMQ_MSG_MORE(flags) (flags & ZMQ_MSG_MORE) ?
[18:36] indygreg LodeRunner: most interesting. I'm not too familiar with the internals. maybe cremes, sustrik, or mikko can help you
[18:36] cremes indygreg: good point; let's see what folks on the ML have to say about it
[18:36] cremes LodeRunner: are you using ipc or inproc transport?
[18:37] LodeRunner cremes: inproc
[18:37] cremes LodeRunner: what version of 0mq?
[18:38] cremes and what os?
[18:38] LodeRunner cremes: currently 2.1.7, linux-x86
[18:38] cremes ok
[18:38] cremes in versions < 2.1.10, the internal commands were passed around using a "mailbox" structure
[18:39] cremes based off of unix sockets
[18:39] cremes in 2.1.10, that was changed
[18:39] cremes any chance you can upgrade and try again?
[18:39] LodeRunner cremes: oh, thanks for the info. time to upgrade, then!
[18:39] LodeRunner sure!
[18:42] cremes let us know if that fixes your issue; it could still be something else
[19:07] LodeRunner cremes: unfortunately, I'm getting the same behavior with 2.1.10
[19:08] LodeRunner on each new thread I'm creating 5 inproc sockets (1 SUB, 2 PUSH, 2 REP); for each new thread I add, lsof counts 12 more descriptors
[19:12] LodeRunner (it may also be relevant that one of the REP sockets is connected to a DEALER in a QUEUE zmq_device)
[19:37] mikko LodeRunner: which os?
[19:38] LodeRunner mikko: linux-x86
[19:38] mikko 12 sounds a bit excessive
[20:02] LodeRunner mikko: my mistake, I'm creating 6 inproc sockets (1 PUB, 1 SUB, 2 PUSH, 2 REP) per thread
[20:02] LodeRunner and there are 12 unix sockets being created, two per zmq_socket (I see the numbers go down as I comment out the zmq_socket's)
[20:15] indygreg LodeRunner: it might help if you post sample code
[20:16] LodeRunner indygreg: ok, I'll try to come up with a sample that isolates the issue
[20:16] mikko LodeRunner: why so many?
[20:17] mikko two fds per inproc socket doesn't sound that bad
[20:17] mikko one for each end
[20:17] mikko i don't know what the new implementation looks like under the hood
[20:23] LodeRunner mikko: you mean 6 zmq_sockets per thread is too many? I've been using them liberally, for various purposes (one to report to the logger thread, one to get tasks from the listener thread, one to get commands from a console thread, etc.)
[20:24] mikko LodeRunner: doesn't large amount of sockets add complexity?
[20:25] LodeRunner mikko: well, before that I was using threads+shared memory+locks to handle all this communication, so converting the code to threads+zeromq+messages simplified things
[20:26] mikko LodeRunner: i mean do you find it easier to handle than one socket per thread and using routing for messages?
[20:26] mikko if you do i guess its fine
[20:26] mikko there should be very little overhead
[22:11] LodeRunner mikko, cremes, indygreg: I just solved my issue of "Too many open files" by raising the value of max_sockets in config.hpp and recompiling zeromq
[22:11] amoffatw hi. i'm trying to understand how pub-sub works. when a subscriber sets its subscribe filter, does this filter register on the publisher side? or does every published message get sent to all subscribers, which filter them on their own?
[22:12] amoffatw the former makes the most sense
[22:12] amoffatw but i just want to be sure
[22:13] minrk amoffatw - depends on what version of zeromq
[22:13] minrk one of the major improvements in zeromq-3 (current beta) is publisher-side filtering
[22:13] amoffatw minrk, ah ok, so if i'm not using the beta, the filtering is subscriber side then
[22:13] minrk correct
[22:13] amoffatw ok that answers it, thanks minrk
[22:15] mikko LodeRunner: how many sockets do you have?
[22:17] cremes LodeRunner: i forgot about that! i always recompile my copy of 0mq to use 51200 sockets. :)
[22:18] LodeRunner mikko: my process now has 1374 sockets, with 100 threads, but that's including a bunch of TCP ports I'm listening as well.
[22:18] mikko LodeRunner: thats a fair amount of sockets
[22:19] LodeRunner cremes: I found this solution in one of the IRC logs, through Google, so I figured I'd better mention the solution of my problem here as well
[22:19] mikko hmm
[22:19] mikko i wonder if max sockets should be configurable in ./configure
[22:19] mikko fairly simple to change
[22:22] LodeRunner mikko: no idea how this number is used internally, but I think the only reason not to make this easy to tweak is if it breaks ABI
[23:21] mbj PULL sockets do not really pull the server? So a busy worker executing a long job cloud easily have many unprocessed messages mailbox that are blocked for a long time?
[23:21] mbj s/cloud/could/
[23:27] minrk yes, in push/pull, the push is really a more accurate descriptor than pull, which only pulls from its local queue
[23:29] mbj minrk: thx for calrification.
[23:30] mbj minrk: Exactly what I understood, but since this is a key reason for one of my design decisions I had to double check this ;)