Assembling solid data in parallel

[k-gun12]

I've got 120 million 50bp SOLiD reads with a 2100bp insert size. Velvet is off the table for assembly as it eats up the 48 gig ram on my workstation pretty quickly. I've got a 160-cpu MPI cluster that should do the trick but the parallel assemblers I'm aware of (abyss, Forge) seem to be mostly untested with ABI data.. does anyone have any more recommendations? I pruned the data down to ~60 million reads and recompiled velvet with a maxk of 21 to use less ram, which runs a little farther but still exhausts system memory. Anyone have other parallel suggestions for assembling my data, or is it time to invest in a giant 512gb machine?

Thanks!

[maasha]

Remove duplicate entries (also reverse-complement). Break the data into 1000 subsets and assemble each and use the resulting contigs as sanger reads with remaining reads (reads not in contigs)?

[k-gun12]

I thought about that, but I'd be losing a fair amount of my paired end data that way, correct? Pairs that aren't placed in a larger contig will just be represented as singletons after each subset is assembled and treated as sanger?

[maasha]

Velvet, and possibly MIRA, will allow you to keep the pair-end info. One problem may be contigs too long to be used as Sanger reads, but those can be shredded into artificial reads in such a way that they will be reconstructed.