[Rd] parallel PSOCK connection latency is greater on Linux?

[Iñaki Ucar]

Please, check a tcpdump session on localhost while running the following script:

library(parallel)
library(tictoc)
cl <- makeCluster(1)
Sys.sleep(1)

for (i in 1:10) {
  tic()
  x <- clusterEvalQ(cl, iris)
  toc()
}

The initialization phase comprises 7 packets. Then, the 1-second sleep
will help you see where the evaluation starts. Each clusterEvalQ
generates 6 packets:

1. main -> worker PSH, ACK 1026 bytes
2. worker -> main ACK 66 bytes
3. worker -> main PSH, ACK 3758 bytes
4. main -> worker ACK 66 bytes
5. worker -> main PSH, ACK 2484 bytes
6. main -> worker ACK 66 bytes

The first two are the command and its ACK, the following are the data
back and their ACKs. In the first 4-5 iterations, I see no delay at
all. Then, in the following iterations, a 40 ms delay starts to happen
between packets 3 and 4, that is: the main process delays the ACK to
the first packet of the incoming result.

So I'd say Nagle is hardly to blame for this. It would be interesting
to see how many packets are generated with TCP_NODELAY on. If there
are still 6 packets, then we are fine. If we suddenly see a gazillion
packets, then TCP_NODELAY does more harm than good. On the other hand,
TCP_QUICKACK would surely solve the issue without any drawback. As
Nagle himself put it once, "set TCP_QUICKACK. If you find a case where
that makes things worse, let me know."

Iñaki

On Wed, 4 Nov 2020 at 04:34, Simon Urbanek <simon.urbanek using r-project.org> wrote:
>
> I'm not sure the user would know ;). This is very system-specific issue just because the Linux network stack behaves so differently from other OSes (for purely historical reasons). That makes it hard to abstract as a "feature" for the R sockets that are supposed to be platform-independent. At least TCP_NODELAY is actually part of POSIX so it is on better footing, and disabling delayed ACK is practically only useful to work around the other side having Nagle on, so I would expect it to be rarely used.
>
> This is essentially RFC since we don't have a mechanism for socket options (well, almost, there is timeout and blocking already...) and I don't think we want to expose low-level details so perhaps one idea would be to add something like delay=NA to socketConnection() in order to not touch (NA), enable (TRUE) or disable (FALSE) TCP_NODELAY. I wonder if there is any other way we could infer the intention of the user to try to choose the right approach...
>
> Cheers,
> Simon
>
>
> > On Nov 3, 2020, at 02:28, Jeff <jeff using vtkellers.com> wrote:
> >
> > Could TCP_NODELAY and TCP_QUICKACK be exposed to the R user so that they might determine what is best for their potentially latency- or throughput-sensitive application?
> >
> > Best,
> > Jeff
> >
> > On Mon, Nov 2, 2020 at 14:05, Iñaki Ucar <iucar using fedoraproject.org> wrote:
> >> On Mon, 2 Nov 2020 at 02:22, Simon Urbanek <simon.urbanek using r-project.org> wrote:
> >>> It looks like R sockets on Linux could do with TCP_NODELAY -- without (status quo):
> >> How many network packets are generated with and without it? If there
> >> are many small writes and thus setting TCP_NODELAY causes many small
> >> packets to be sent, it might make more sense to set TCP_QUICKACK
> >> instead.
> >> Iñaki
> >>> Unit: microseconds
> >>>                    expr      min       lq     mean  median       uq      max
> >>>  clusterEvalQ(cl, iris) 1449.997 43991.99 43975.21 43997.1 44001.91 48027.83
> >>>  neval
> >>>   1000
> >>> exactly the same machine + R but with TCP_NODELAY enabled in R_SockConnect():
> >>> Unit: microseconds
> >>>                    expr     min     lq     mean  median      uq      max neval
> >>>  clusterEvalQ(cl, iris) 156.125 166.41 180.8806 170.247 174.298 5322.234  1000
> >>> Cheers,
> >>> Simon
> >>> > On 2/11/2020, at 3:39 AM, Jeff <jeff using vtkellers.com> wrote:
> >>> >
> >>> > I'm exploring latency overhead of parallel PSOCK workers and noticed that serializing/unserializing data back to the main R session is significantly slower on Linux than it is on Windows/MacOS with similar hardware. Is there a reason for this difference and is there a way to avoid the apparent additional Linux overhead?
> >>> >
> >>> > I attempted to isolate the behavior with a test that simply returns an existing object from the worker back to the main R session.
> >>> >
> >>> > library(parallel)
> >>> > library(microbenchmark)
> >>> > gcinfo(TRUE)
> >>> > cl <- makeCluster(1)
> >>> > (x <- microbenchmark(clusterEvalQ(cl, iris), times = 1000, unit = "us"))
> >>> > plot(x$time, ylab = "microseconds")
> >>> > head(x$time, n = 10)
> >>> >
> >>> > On Windows/MacOS, the test runs in 300-500 microseconds depending on hardware. A few of the 1000 runs are an order of magnitude slower but this can probably be attributed to garbage collection on the worker.
> >>> >
> >>> > On Linux, the first 5 or so executions run at comparable speeds but all subsequent executions are two orders of magnitude slower (~40 milliseconds).
> >>> >
> >>> > I see this behavior across various platforms and hardware combinations:
> >>> >
> >>> > Ubuntu 18.04 (Intel Xeon Platinum 8259CL)
> >>> > Linux Mint 19.3 (AMD Ryzen 7 1800X)
> >>> > Linux Mint 20 (AMD Ryzen 7 3700X)
> >>> > Windows 10 (AMD Ryzen 7 4800H)
> >>> > MacOS 10.15.7 (Intel Core i7-8850H)
> >>> >
> >>> > ______________________________________________
> >>> > R-devel using r-project.org mailing list
> >>> > https://stat.ethz.ch/mailman/listinfo/r-devel
> >>> >
> >>> ______________________________________________
> >>> R-devel using r-project.org mailing list
> >>> https://stat.ethz.ch/mailman/listinfo/r-devel
> >> --
> >> Iñaki Úcar
> >
> > ______________________________________________
> > R-devel using r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-devel
> >
>


-- 
Iñaki Úcar