Skip to content

Using log() on an openMosix cluster

2 messages · Roger D. Peng, Torsten Hothorn

Original
Roger D. Peng
# 
Hi all, I was hoping to get some advice about a problem that I realize 
will be difficult to reproduce for some people.  I'm running R 1.7.1 on 
an openMosix (Linux) cluster and have been experiencing some odd 
slow-downs.  If anyone has experience with such a setup (or a similar 
one) I'd appreciate any help.  Here's a simplified version of the problem.

I'm trying to run the following code:
##
N <- 100000; a <- numeric(N); b <- numeric(N)
e <- rnorm(N)

for(i in 1:N) {
         a[i] <- exp(e[i])
         b[i] <- log(abs(a[i]))
}
##

When I run it on the head node, everything is fine.  However, when I 
send the R process off to one of the cluster nodes (i.e. using mosrun 
from the head node) the program takes about 10 times longer (in 
wall-clock time, cpu time is roughly the same).

Interestingly, when I tried running the following code:
##
N <- 100000; a <- numeric(N); b <- numeric(N)
e <- rnorm(N)

for(i in 1:N) {
         a[i] <- exp(e[i])
         b[i] <- exp(abs(a[i]))
}
##

I didn't experience any slow-down!  That is the wall-clock time is the 
same when run on the head node or on the cluster nodes.  The only 
difference between the two programs is that one takes a log in the for() 
loop and the other one takes an exponential.

I guess my question is why would taking the log() produce a 10 fold 
increase in runtime?  I know that Mosix clusters can experience serious 
performance hits if you make a lot of system calls or write out data to 
files but I don't think I'm doing that here.  Is there some major 
difference in the way that exp() and log() are implemented?

I'm pretty sure this isn't an R problem but I'm wondering if R is doing 
something behind the scenes that's affecting performance in the 
openMosix setting.

Thanks in advance for any help.

-roger
2 days later
Torsten Hothorn
#
On Fri, 21 Nov 2003, Roger D. Peng wrote:

            

      
      
      
    

        
Did you adapt the sig*jmp definitions in src/include/Defn.h? This was
necessary until R-1.7.1 and is no longer needed, thanks to Luke's changes
in 1.8.0:

    o   On Unix-like systems interrupt signals now set a flag that is
        checked periodically rather than calling longjmp from the
        signal handler.  This is analogous to the behavior on Windows.
        This reduces responsiveness to interrupts but prevents bugs
        caused by interrupting computations in a way that leaves the
        system in an inconsistent state.  It also reduces the number
        of system calls, which can speed up computations on some
        platforms and make R more usable with systems like Mosix.


I tried the example above with N = 1.000.000:

N <- 1000000; a <- numeric(N); b <- numeric(N)
e <- rnorm(N)

for(i in 1:N) {
         a[i] <- exp(e[i])
         b[i] <- log(abs(a[i]))
}

cat(proc.time())

with R-1.8.0 with Linux 2.4.22 and OpenMosix-Patch and started 10
processes which migrated immediately.

hothorn@mosix:~/tmp/log$ grep -1 cat *.Rout
log1.Rout-R>
log1.Rout:R> cat(proc.time())
log1.Rout-37.04 1.02 43.44 0 0.01R>
--
log10.Rout-R>
log10.Rout:R> cat(proc.time())
log10.Rout-34.25 0.45 40.21 0 0.01R>
--
log2.Rout-R>
log2.Rout:R> cat(proc.time())
log2.Rout-22.19 0.33 29.36 0 0R>
--
log3.Rout-R>
log3.Rout:R> cat(proc.time())
log3.Rout-24.46 0.42 32.96 0 0.03R>
--
log4.Rout-R>
log4.Rout:R> cat(proc.time())
log4.Rout-36.88 0.38 40.73 0 0.02R>
--
log5.Rout-R>
log5.Rout:R> cat(proc.time())
log5.Rout-34.79 0.52 42.83 0.02 0R>
--
log6.Rout-R>
log6.Rout:R> cat(proc.time())
log6.Rout-34.14 0.54 41.46 0 0.01R>
--
log7.Rout-R>
log7.Rout:R> cat(proc.time())
log7.Rout-35.21 0.66 43.4 0 0R>
--
log8.Rout-R>
log8.Rout:R> cat(proc.time())
log8.Rout-25.27 0.55 33.77 0 0.01R>
--
log9.Rout-R>
log9.Rout:R> cat(proc.time())
log9.Rout-36.69 0.44 43.16 0.01 0R>

So, everything is fine here. I guess using R-1.8.1 will fix your
problem.

Torsten