Re: [go-nuts] Multiplexing blocking call on Go routine

[zohaibsh]

I think there a confusion of what I am trying to say. With the go routine 
approach there is cost of 1M connections + cost of 1M go routines running a 
blocking call all the time. With an event based system where I can 
multiplex code execution over a fixed size go routines, the cost is already 
cut in half because I can do 1M connections + cost of 1K go routines. But I 
get the gist, there is no way for me except to go really go on something 
like libuv.

On Wednesday, 4 January 2017 19:14:32 UTC-8, Jesse McNelis wrote:
>
> On Thu, Jan 5, 2017 at 9:51 AM,   wrote: 
> > Hey guys, 
> > 
> >  So for some time now I have been trying to build a high performance 
> Pub/Sub 
> > server in Go. I am using gorilla websocket library, which in it self has 
> a 
> > Read, Write methods 
> > (https://godoc.org/github.com/gorilla/websocket#Conn.ReadMessage) that 
> will 
> > block my current routine. In order to continuously listen to messages 
> coming 
> > from client I spin off a go routine, reading from socket and pushing 
> > messages over a channel. Which I in-turn can use for doing select { } in 
> > another go routine. When I testing this thing for scaling, I just wrote 
> some 
> > test code and turns out spinning off 1M go routines (doing nothing) 
> results 
> > in ~8GB of memory usage on my windows machine. 
>
> You should compare this with a simple C program using epoll, you'll 
> find that maintaining a 1M connections uses quite a lot of memory. 
>
> >These go routines are doing 
> > nothing (sleeping) and it requires this much memory; which forced me to 
> ask 
> > if I can multiplex such blocking calls on a single go routine. If I can 
> do 
> > so I can spin off something like 1000 go routines and each one of them 
> > multiplexing 1000 connections (which in theory should require lesser 
> memory 
> > that what I am currently getting). 
>
> But a real program won't holding 1M connections and doing nothing with 
> them. 
> The cost of processing requests from those connections will be much 
> larger than the cost of a goroutine. 
>
>
> >  Any ideas how can one call muiltple socket reads (something like event 
> > loop) on single go routine. 
>
> There really isn't a way, The net package already provides a 
> lightweight way to maintain larger numbers of connections using 
> goroutines. 
>
>
> > Disclaimer: I know event loop may sound against the golang philosophy 
> but I 
> > see no reason to pay such huge memory cost when most of them will be 
> idle, 
> > waiting for something to come from user. 
>
> I think you're underestimating the memory cost of simply having 1M 
> connections. 
>

-- 
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.
For more options, visit https://groups.google.com/d/optout.

Re: [go-nuts] Multiplexing blocking call on Go routine

[Jesse McNelis]

On Thu, Jan 5, 2017 at 9:51 AM,   wrote:
> Hey guys,
>
>  So for some time now I have been trying to build a high performance Pub/Sub
> server in Go. I am using gorilla websocket library, which in it self has a
> Read, Write methods
> (https://godoc.org/github.com/gorilla/websocket#Conn.ReadMessage) that will
> block my current routine. In order to continuously listen to messages coming
> from client I spin off a go routine, reading from socket and pushing
> messages over a channel. Which I in-turn can use for doing select { } in
> another go routine. When I testing this thing for scaling, I just wrote some
> test code and turns out spinning off 1M go routines (doing nothing) results
> in ~8GB of memory usage on my windows machine.

You should compare this with a simple C program using epoll, you'll
find that maintaining a 1M connections uses quite a lot of memory.

>These go routines are doing
> nothing (sleeping) and it requires this much memory; which forced me to ask
> if I can multiplex such blocking calls on a single go routine. If I can do
> so I can spin off something like 1000 go routines and each one of them
> multiplexing 1000 connections (which in theory should require lesser memory
> that what I am currently getting).

But a real program won't holding 1M connections and doing nothing with them.
The cost of processing requests from those connections will be much
larger than the cost of a goroutine.


>  Any ideas how can one call muiltple socket reads (something like event
> loop) on single go routine.

There really isn't a way, The net package already provides a
lightweight way to maintain larger numbers of connections using
goroutines.


> Disclaimer: I know event loop may sound against the golang philosophy but I
> see no reason to pay such huge memory cost when most of them will be idle,
> waiting for something to come from user.

I think you're underestimating the memory cost of simply having 1M connections.

-- 
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.
For more options, visit https://groups.google.com/d/optout.