Profiling shows ~150_000 CS for channel version. For some reason I thought
scheduling should be done by Go runtime, but it is not.
Anyway you are right, I will look closer into syscall part, may be I will
get rid of netlink library, or something..
On Monday, February 4, 2019 at 6:59:33 PM UT
And if that is the case, the most likely way to speed up you program would be
to send the route updates in a batch to the kernel - not sure that this is even
possible.
> On Feb 4, 2019, at 9:56 AM, robert engels wrote:
>
> Also, to clarify, in the absence of more details, I would consider
> n
Also, to clarify, in the absence of more details, I would consider
netlink.RouteReplace to be an IO operation - it is calling into the kernel
(most likely). Whether that should be considered IO is probably a matter of
opinion.
> On Feb 4, 2019, at 9:32 AM, Andrew Medvedev
> wrote:
>
> Hi Rob
That doesn’t seem to match up with the pprof report - although it would be
easier to determine if you showed the back-traced the syscalls. The original
pprof report you stated was from without channels, thus the syscalls are almost
certainly related to IO. Maybe you could clarify why you think i
Hi Robert
The program is not IO bound. Sorry for not giving out the source code in
the first place, here is a no-channel
version https://github.com/andrewmed/loadroutes/blob/master/cmd/main.go and
here is a slower channels
version https://github.com/andrewmed/loadroutes/blob/test/channels/cmd/
Reading and writing to a channel has locks behind the scenes. Locks are
implemented with futexes.
If your program is IO bound the simplest solution is often to use buffering in
the IO to perform more work per IO operation.
> On Feb 3, 2019, at 5:00 PM, Andrew Medvedev
> wrote:
>
> Hi every
Hi everybody
I have a simple program which parses IP addresses and adds them to the
system routing table using the netlink library. The number is routes is
huge (400_000). My parsing part is very simple and fast, and most of the
time is spent in system calls:
File: loadroutes
Type: cpu
T
