TTK Ciar's Journal: More Minor Jibber-Jabber

2006-11-14: More Minor Jibber-Jabber

These Are Minor Journal Entries so don't expect too much from them, 'k?

2006-11-14: That Went Well, and Hey PVM!

My previous major journal entry, Here Comes The Old Boss, Same As The Old Boss, was well-received for all that it was hastily-pounded-out, mostly-unsupported crap! :-) I felt badly for not backing up my position with more concrete evidence, but people seem to like it anyway. Guess I'll make more!

A friend of mine who stays more informed than I do about the details of individual politicians' activities says that I'm a little too harsh on our political leaders. I've told him that I'll keep an open mind, and would like to document examples he comes up with of politicians who display noncorrupt, responsible behavior. I'd like my perspectives to be fair.

In other news, I've been working lately to like MPI, the "new" message-passing interface which has largely replaced PVM. I've been working really hard to like it. MPI makes liking it extremely difficult. It is a bloated, needlessly convoluted system with some odd gaps in its capabilities. I get the distinct impression that PVM was developed by engineers to get real-world work done, whereas MPI was developed by academicians to get their names on little pieces of paper.

Today I broke with my discipline and wrote some code for work which uses PVM. It felt so good! The coding was fun and easy, debugging was short, and the (admittedly simple) application works like a dream. I went from zero to fully-functioning service in about four hours. I whugglez my PVM.

Nevertheless, I really should learn MPI. Why? Well, a few reasons. First, nobody's hiring distributed system engineers for their PVM skills. Everyone wants MPI. At The Archive I'm left to my own devices, but I might not be there forever, so getting some MPI experience under my belt could improve my chances of working with distributed systems for my next employer. Aside from that, PVM has some shortcomings which MPI addresses. MPI passes messages directly between individual processes, which incurs less overhead than PVM's process->master->process relayed path (or process->master->master->process when communicating between servers). Also, PVM is intrinsically limited to 4096 nodes per virtual machine, is vulnerable to decapitation (the PVM "master node" represents a single point of failure which cannot be failed-over), and provides only a very limited means of incorporating new nodes into the virtual machine (rsh or ssh connections made from the master node to the new node, over which pvmd handshaking is negotiated).

I would like to work with a virtual machine framework that lacks these shortcomings, and so far my best idea is to implement one on top of MPI (which is just a message-passing system, and not a virtual machine like PVM). If MPI proves too annoying, I'll implement my own message-passing system, but would rather avoid that. (Well, sort of. I'd love to implement a nice message-passing system, but have limited time and too many other projects, and should put MPI on my resume too.)

Okay, enough blather for today.

-- TTK