[Work Log] Dis Prop (ctd); new HPC

Continue writing dissertation proposal. Worked on SfM, MVS literature review. Spent quite some time trying to understand self-calibration. The issue is that with a general perspective camera, SfM can only perform reconstruction up to a projective ambiguity (most of the SfM literature of late is apparently reticent about this?). It turns out, if we fix at least one intrinsic parameter (how about skew = 0? or aspect ration = 1?) then we get a metric reconstruction (which maybe explains why it's no longer covered in SfM papers).

This shows that bundle adjustment can give a metric reconstruction, but the question remains how to initialize it. Pollefeys et. al (1998) propose an analytical self-calibration solution. Snavely et al. (2007) apparently just use bundle adjustment anyway, presumably using a simplified pinhole camera (no skew or ppo) with a reasonable default for f, or EXIF tags when possible. Brown and Lowe (2005) initialize each new camera with the old camera's intrinsics, no word on how the first pair is initialized (maybe obvious but I missed it?).

New HPC

UA just installed their new 80-node HPC cluster and by a stroke of luck, Ive gotten early access to it! It looks like all machines have 16 CPU cores and 256 GB RAM; 60 nodes have multiple GPU cards while the other 20 have Intel PHI general purpose compute cards. It's amazing to query the cluster and see the entire thing at 0% utilization... that won't last long! Having fun poking around it and reading the IBM LSF manuals 2 (the job queueing system).

I found the following resource to be very useful, even though it's from University of Miami's personal LSF installation:

• unofficial LSF 9.1.1 Documentation site (University of maimi)
  • Basics
  • Jobs

The hew official documents I could find were less useful:

• Command Reference [pdf]
• Administration Reference [pdf]
• Configuration Reference[pdf]

Notes

Okay great! Got a general idea of how to use the system from the user guide

• bsub - submit default job queue
• bsub -q <queue-name> - submit specific job queue
• bsub -I - interactive job
• bjobs - view submitted jobs
• bqueues - list info for all queues
• bparams - details for default queue
• lsload - list load on hosts
• lshosts - list hosts w/ configuration
• bhosts - view batch server hosts (what's a batch server?)
• lsid - cluster info (name and master host)
• lsrun - run a command on a free host
• lsgrun - run a command on a group of free hosts
• lsltakss - view and add local tasks (huh?)
• lsrtasks - view and add remote tasks (huh?)
• bkill 1234 - kill job 1234
• bkill -r - force removal of job (if hung after kill?)
• bstop - suspend job
• `bresume - resume job
• btop 1234 - move job 1234 to top of queue
• bbot 1234 - move job 1234 to bottom of queue

On gpu* machines

• nvidia-smi list details for all gpu cards (and running processes?)

On phi* machines

• micinfo lists details for all coprocessor cards

Also MPI is supported

Discovered two large gsfs disks (192 TB and 43 TB, resp.). Apparently GSFS is a GPU-enabled encrypted disk. No write permissions ATM

Questions

Some random thoughts while using the system

• can/should we get Matlab? Octave?
• Subversion?
• write permissions on GSFS disks?
• No job report email after calling bsub -u ksimek@email.arizona.edu

Posted by Kyle Simek

blog comments powered by Disqus