Re: [igraph] page.rank speed

[Sam Steingold]

Tamas, thanks for your reply.
the calculation just died:

Error in page.rank(graph) :
  At arpack.c:1178 : ARPACK error, The Schur form computed by LAPACK
routine dlahqr could not be reordered by LAPACK routine dtrsen.
Timing stopped at: 127517.2 3.297 127803.5

I tried

page.rank(graph, options=list(niter=2))

and it is still running:
  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
36756 sam       20   0 13.8g  13g 4564 R 99.6 43.5   1131:50
/usr/lib64/R/bin/exec/R --no-save --no-restore -q

the help page for arpack which describes options for page.rank does
not mention epsilon.

On Wed, Jul 31, 2013 at 4:23 PM, Tamás Nepusz <address@hidden> wrote:
>> page.rank has been running forever:
> [...]
>> on
>>
>> IGRAPH DN-- 12062801 15279166 --
>> attr: name (v/c)
>>
>> is this normal?
> Not sure but I wouldn't rule it out -- your graph is pretty large and ARPACK 
> can be a tad bit slow. I have just tried it on an Erdos-Renyi graph of 
> similar size and it also did not terminate for a while. On the other hand, 
> you might be experiencing a convergence issue with ARPACK. Sometimes it 
> happens that ARPACK fails to detect convergence (at least within the machine 
> epsilon) and performs the maximum number of allowed iterations with 
> essentially no improvement after the first few iterations -- then it bails 
> out with a warning saying that the maximum number of iterations has been 
> reached. Try increasing the epsilon value a bit when invoking page.rank() to 
> see if it terminates faster or not. Or decrease the maximum number of 
> iterations allowed. I'm not familiar with the R interface but I think there 
> is an options=... argument to page.rank where you can specify some options 
> that ARPACK will use.
>
> An alternative solution if you are feeling adventurous and you don't mind 
> compiling igraph from source: try the 0.7-prpack branch of igraph 
> (downloadable from Github):
>
> https://github.com/igraph/igraph/tree/0.7-prpack
>
> This branch uses the excellent PRPACK package instead of ARPACK to calculate 
> PageRank scores. PRPACK does all sorts of neat tricks to ensure faster 
> convergence for PageRank calculations. The branch will be merged with the 0.7 
> development tree sooner or later but it still needs more testing.
>
> All the best,
> Tamas



-- 
Sam Steingold <http://sds.podval.org> <http://www.childpsy.net/>