On Thursday, March 15, 2018 at 4:23:41 AM UTC+8, Bakul Shah wrote: > > > You will need 10K*64KiB*2 = 1.22GiB kernel bufferspace at a > minimum in any case. That is still one datagram's worth of > buffering per UDP socket so you can still lose some packets. > This is in addition to any user level buffering passd in to > recv() or read(). So you need at least 2.48GiB! >
Yeah, there's no way to avoid the kernel buffer waste. For the UDP socket I'm fine losing a few packets from time to time because it does not promise reliable delivery anyway. > I did such a custom poller for a network proxy. It was a pain > to get right (I only did the Eww!poll version). In my case all > the traffic was via tcp. It has been a while now but IIRC my > reason was not to save memory but to gain flexibility over Go > runtime's networking. I don't remember the details now though. > I tried and it was indeed a lot of work. So I'm trying to see if there's a middle ground that could keep using goroutine (2kb stack is acceptable) but avoid the data buffer waste. Please see my previous reply about exposing a tiny bit of information from the internal poller to address the issue. I'd like to see if there's any downside of that. Thanks! :) -- 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.
