I meant lock-free as in "without explicit locks". The original challenge still stands if someone has a better solution than me: "The deadlocks in 2_1.go and 2_2.go are caused by the simplistic and wrong implementation of bidi-comm, which is what I'll be illustrating. I have three working solutions - 1_1.go, 2_3.go, 2_4.go. So the question is, can we remove the extra goroutine from 1_1.go and make the code nicer to read than 2_3.go and 2_4.go. The extra goroutine that I'd like to be removed is started here: https://github.com/egonk/chandemo/blob/master/1_1.go#L14 (line 14)"
On Sunday, December 8, 2019 at 7:18:16 AM UTC+1, Robert Engels wrote: > > I understand what you are saying but I’ll still suggest that your > premise/design is not correct. There are plenty of useful lock free > structures in Go (see github.com/robaho/go-concurrency-test) but that is > not what you are attempting here... you are using async processing - these > are completely different things. Using async in Go is an anti-pattern IMO. > > On Dec 8, 2019, at 12:11 AM, Egon Kocjan <eko...@gmail.com <javascript:>> > wrote: > > > I'll cite myself: > "I'm preparing a short talk about Go channels and select. More > specifically, I want to show what not to do." > and > "it would be tempting to just combine two goroutines into one and handle > caching in a single loop without using locks (I see developers avoid > atomics and locks if they don't have a lot of previous experience with > traditional MT primitives)" > > Before I say one can't do something in Go, I wanted to ask here to make > sure I'm not missing something obvious. Basically, I intend to show how > difficult lock-free programming can be so don't force it - just use > goroutines and locks. > > On Saturday, December 7, 2019 at 3:46:43 PM UTC+1, Robert Engels wrote: >> >> Probably not. Go is designed for 1:1 and there is no reason to do it >> differently. You could probably try to write an async event driven layer >> (which it looks like you’ve tried) but why??? >> >> It’s like saying I’d really like my plane to float - you can do that -but >> most likely you want a boat instead of a plane. >> >> On Dec 7, 2019, at 2:38 AM, Egon Kocjan <eko...@gmail.com> wrote: >> >> >> I'll try to clarify as best as I can, thanks again to anyone looking at >> this. >> >> The simple server implementation of "output <- input+1" is here and it is >> not "under our control" - it's what we have to work with: >> https://github.com/egonk/chandemo/blob/master/server.go >> >> The test runner or client is here: >> https://github.com/egonk/chandemo/blob/master/demo.go (it just pushes in >> ints and gets server replies back through a connection layer) >> >> The deadlocks in 2_1.go and 2_2.go are caused by the simplistic and wrong >> implementation of bidi-comm, which is what I'll be illustrating. I have >> three working solutions - 1_1.go, 2_3.go, 2_4.go. So the question is, can >> we remove the extra goroutine from 1_1.go and make the code nicer to read >> than 2_3.go and 2_4.go. The extra goroutine that I'd like to be removed is >> started here: >> https://github.com/egonk/chandemo/blob/master/1_1.go#L14 (line 14) >> >> What I mean by removed - no go statement, replaced presumably by some >> kind of for/select combination. >> >> On Saturday, December 7, 2019 at 7:02:50 AM UTC+1, robert engels wrote: >>> >>> I’m sorry but your design is not comprehendible by me, and I’ve done >>> lots of TCP based services. >>> >>> i think you only need to emulate classic TCP processing - a reader >>> thread (Go routine) on each side of the connection using range to read >>> until closed. The connection is represented by 2 channels - one for each >>> direction. >>> >>> I think you might be encountering a deadlock because the producer on one >>> end is not also reading the incoming - so either restructure, or use 2 more >>> threads for the producers. >>> >>> >>> >>> On Dec 6, 2019, at 10:38 PM, Egon Kocjan <eko...@gmail.com> wrote: >>> >>> Agreed, I see goroutines in general as a big win. But what I intend to >>> talk about in the presentation: >>> - we have two unidirectional flows of data resembling something like a >>> TCP socket, easy to do with two goroutines with a for loop >>> - let's add caching, so some requests do not go to the server >>> - it would be tempting to just combine two goroutines into one and >>> handle caching in a single loop without using locks (I see developers avoid >>> atomics and locks if they don't have a lot of previous experience with >>> traditional MT primitives) >>> - this is surprisingly difficult to do properly with Go channels, see my >>> attempts: https://github.com/egonk/chandemo/blob/master/2_3.go and >>> https://github.com/egonk/chandemo/blob/master/2_4.go >>> <https://github.com/egonk/chandemo/blob/master/2_3.go> >>> - it is easy to do in actor systems, just move the code for both actors >>> into a single actor! >>> >>> The lesson here is that select is not a nice and safe compose statement >>> even if it appears so at the first glance, do not be afraid to use locks. >>> >>> Of course, if somebody comes up with a better implementation than 2_3.go >>> and 2_4.go, I would be very happy to include it in the talk. >>> >>> On Saturday, December 7, 2019 at 4:17:04 AM UTC+1, robert engels wrote: >>>> >>>> To clarify, with Go’s very lightweight threads it is “doing the >>>> multiplexing for you” - often only a single CPU is consumed if the >>>> producer >>>> and consumer work cannot be parallelized, otherwise you get this >>>> concurrency “for free”. >>>> >>>> You are trying to manually perform the multiplexing - you need async >>>> structures to do this well - Go doesn’t really support async by design - >>>> and it’s a much simpler programming model as a result. >>>> >>>> On Dec 6, 2019, at 12:02 PM, Robert Engels <ren...@ix.netcom.com> >>>> wrote: >>>> >>>> A channel is much closer to a pipe. There are producers and consumers >>>> and these are typically different threads of execution unless you have an >>>> event based (async) system - that is not Go. >>>> >>>> On Dec 6, 2019, at 9:30 AM, Egon Kocjan <eko...@gmail.com> wrote: >>>> >>>> >>>> There are goroutines in the examples of course, just a single goroutine >>>> per bidi channel seems hard. By contrast, I've worked with actor systems >>>> before and they are perfectly fine with a single fiber. >>>> >>>> On Friday, December 6, 2019 at 3:38:20 PM UTC+1, Robert Engels wrote: >>>>> >>>>> Channels are designed to be used with multiple go routines - if you’re >>>>> not you are doing something wrong. >>>>> >>>>> On Dec 6, 2019, at 8:32 AM, Egon Kocjan <eko...@gmail.com> wrote: >>>>> >>>>> >>>>> Hello >>>>> >>>>> I'm preparing a short talk about Go channels and select. More >>>>> specifically, I want to show what not to do. I chose a bidirectional >>>>> communication channel implementation, because it seems to be a common >>>>> base >>>>> for a lot of problems but hard to implement correctly without using any >>>>> extra goroutines. All the code is here: >>>>> https://github.com/egonk/chandemo >>>>> >>>>> 1_1.go: easy with en extra goroutine (takes 1.2s for million ints) >>>>> 2_1.go: nice but completely wrong >>>>> 2_2.go: better but still deadlocks >>>>> 2_3.go: correct but ugly and slow (takes more than 2s for million ints) >>>>> 2_4.go: correct and a bit faster but still ugly (1.8s for million ints) >>>>> >>>>> So my question: is there a better way of doing it with just nested for >>>>> and select and no goroutines? Basically, what would 2_5.go look like? >>>>> >>>>> Thank you >>>>> Egon >>>>> >>>>> -- >>>>> You received this message because you are subscribed to the Google >>>>> Groups "golang-nuts" group. >>>>> To unsubscribe from this group and stop receiving emails from it, send >>>>> an email to golan...@googlegroups.com. >>>>> To view this discussion on the web visit >>>>> https://groups.google.com/d/msgid/golang-nuts/82830a5d-2bd8-4324-890e-9ae7f5f0fbaf%40googlegroups.com >>>>> >>>>> <https://groups.google.com/d/msgid/golang-nuts/82830a5d-2bd8-4324-890e-9ae7f5f0fbaf%40googlegroups.com?utm_medium=email&utm_source=footer> >>>>> . >>>>> >>>>> >>>> -- >>>> You received this message because you are subscribed to the Google >>>> Groups "golang-nuts" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to golan...@googlegroups.com. >>>> To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/golang-nuts/bdc57eb0-b26f-4364-87fb-241b0807e8ae%40googlegroups.com >>>> >>>> <https://groups.google.com/d/msgid/golang-nuts/bdc57eb0-b26f-4364-87fb-241b0807e8ae%40googlegroups.com?utm_medium=email&utm_source=footer> >>>> . >>>> >>>> >>>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "golang-nuts" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to golan...@googlegroups.com. >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/golang-nuts/75d69b4e-4fb7-4f62-8011-f21e2a4c294a%40googlegroups.com >>> >>> <https://groups.google.com/d/msgid/golang-nuts/75d69b4e-4fb7-4f62-8011-f21e2a4c294a%40googlegroups.com?utm_medium=email&utm_source=footer> >>> . >>> >>> >>> -- >> You received this message because you are subscribed to the Google Groups >> "golang-nuts" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to golan...@googlegroups.com. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/golang-nuts/8b87adcc-2249-402c-b34c-20df5013860a%40googlegroups.com >> >> <https://groups.google.com/d/msgid/golang-nuts/8b87adcc-2249-402c-b34c-20df5013860a%40googlegroups.com?utm_medium=email&utm_source=footer> >> . >> >> -- > You received this message because you are subscribed to the Google Groups > "golang-nuts" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to golan...@googlegroups.com <javascript:>. > To view this discussion on the web visit > https://groups.google.com/d/msgid/golang-nuts/3b9bb722-d43f-4e70-8384-dc17cdec6090%40googlegroups.com > > <https://groups.google.com/d/msgid/golang-nuts/3b9bb722-d43f-4e70-8384-dc17cdec6090%40googlegroups.com?utm_medium=email&utm_source=footer> > . > > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/4a176af0-74bb-49b5-ae4d-d8714c7bc46d%40googlegroups.com.
