On 03/01/2013 10:55 PM, Louis Letourneau wrote:
> I know I ask this often, but I don't grasp yet when something is a
> possible bug (like the infinite repeat loop) or just normally long.
>
> If I run an assembly with 2000 cores and run the same with 1000 cores
> (say I have still enough ram per cores) should is be close to linearly
> slower?
>

Most of the step will scale almost linearly if the latency remains constant.
The extension of seeds, however, scale, but the scaling slope is less prominent.

> I ran an assembly awhile back with k61, 2025 cores with ray 2.1.0 +
> patch5361, it took 47 hrs.
>

If you are using a kmer length at least 51, you can add -minimum-contig-length 
115
to reduce the number of seeds. There is some work in progress to fix this at 
the core
too so you don't have to.

> I'm running practically the same with Ray 2.2.0-rc0 with 1024 cores and
> it's been running for 106 hours now.
>
> Last lines in the log:
> Rank 370 starts on seed 4220, length is 92, flow 0 [4220/8071]
> Rank 370 starts on seed 4221, length is 92, flow 0 [4221/8071]
>
>

Length 92 -- this is one short seed that is probably just a tip in the graph.
To avoid these, the trick right now is to use -minimum-contig-length xyz
where xyz is at least 2*k+1.

>
> No file except stdout had changed in the last 3days, and this is what
> worries me a bit more.
>

The verbosity of the program was reduced to reduced the size of the
standard output file.

> Any tricks, hints at what to look for (ElapsedTime hasn't changed either)
>

If you are on a super computer that allows that, you can ssh onto one of the 
node and
attach to one of the processes with


   gdb attach 1234

then run

   backtrace


Here, 1234 is the process identifier.

> Thanks
> Louis
>

Right now, I am finishing a few features in Ray Cloud Browser. One of these 
features will
help me trouble shoot some of the remaining issues for the milestone of Ray 
v2.2.0, which is
past due by 2 days.

  
> ------------------------------------------------------------------------------
> Everyone hates slow websites. So do we.
> Make your web apps faster with AppDynamics
> Download AppDynamics Lite for free today:
> http://p.sf.net/sfu/appdyn_d2d_feb
> _______________________________________________
> Denovoassembler-users mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/denovoassembler-users
>


------------------------------------------------------------------------------
Everyone hates slow websites. So do we.
Make your web apps faster with AppDynamics
Download AppDynamics Lite for free today:
http://p.sf.net/sfu/appdyn_d2d_feb
_______________________________________________
Denovoassembler-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/denovoassembler-users

Reply via email to