Academia con Queso part 1: THE TREADMARKS DISTRIBUTED SHARED MEMORY SYSTEM

why ya checkin out those?

funny to say that a "shared memory interface" is better than a "message passing interface" when that probably uses various message passing interfaces down below to work.

i'm surprised they aren't just acknowledging that their interface is higher level and easier to work with, etc etc.

or something.

what do i know?
m.

The difference is that the synchronization is implicit here -- pages are protected using the mprotect() system call. Then the VM traps access to those and the signal handlers arrange for the pages to be synchronized/replicated etc.

The multiple writer aspect is pretty interesting since it transmits only binary diffs between versions of pages on sync events. This eliminates page contention through false sharing. Example: Proc 1 is writing to memory location a and Proc 2 to (a+1). Both locations are on the same shared page. By using diffs, multiple processors can access discrete objects safely+fast even if they're on a single page.

i was trolling probably. networking = messages, etc., which is a stupid thing to point out.

in college i got to play with the maspar some... well, the diff was that the maspar was under the hood from our point of view. there wasn't anything different than using it on a single proc.

The MasPar MP-2 comes in sizes of 1024, 2048, 4096, 8192 or 16384 processors in its processor element (PE) array. The MP-2 at UTK/JICS has 4096 processors, and we will refer to this machine from here on, unless specifically stated otherwise. A 12-MIPS R (ACU) controls the PE array; it has its own memory for data and instructions.

Every PE receives the same instruction from the ACU simultaneously. Only those processor elements located in the active set, which can be user-defined, carry out instructions received from the ACU. Each of the PEs are 32-bit ALUs under the control of the ACU, with 64 KB of memory and sixty-four 32-bit registers. Both the ACU and the PE array are contained in the DPU, or Data Parallel Unit, which is where all parallel processing is performed. The front end of the MP-2 is a DEC 5000/200 workstation, currently running under Ultrix v4.2a. This front end provides user access to the ACU.

our front end was an sg irix i think. computer vision. missles. blah. that's a drastically different setup tho... cause the maspar is specialized architecture.
m.

jeeez! I only have access to a 24 processor (with 2 chipz on die) Power4 computer and 8 and 16 proc SunFire machines.

Although I might evaluate the Cashmere 2L DSM system on a linux cluster with a MicroChannel bus between nodes. (Cashmere is like TreadMarks with some important differences; the 2L version of Cashmere is SMP-smart)

should i make my presentation on distributed memory have anthropomorphic memory pages ala the Bill in School House Rock?

so is your presentation on that system or are you involved with that system?

either way, that's rulin'.

i like the anthropomorphic memory page idea. there should be more of that in computer slides. (as long as you know you're audience can take a joke and/or handle comic illustrations.)

we had some cool parallel jammers at my undergrad school. before i left we got this ibm supercomputer that was 5 rsc32000's i think... forgetting... they were connected via a superfast dedicated lan. one of the machines was the control box where you staged your stuff... i think some version of load leveller handled the division of work. i wonder if they ever used that? it's funny how these goobs write grant proposals and then sit on hardware like that. no, actually, it's embarassing and sad. what can ya do tho? i think they were gonna datamine the interweb with it.

seems like a fat disc array was supposed to come for it. i dunno. i used to work for the department and that's why i knew about all the special machines we had sitting around. even the maspar never got played with much after certain research was done/completed/fizzled/whatever. like i mentioned the missle crap, i think research for it moved to an air force base after they got so far. i just had a class on computer image processing that used it for kicks and giggles. i woulda studied computer vision in college but got disenchanted when i learned that most of the money for that was coming from the air force.

i mean, it's unavoidable that technologies could be used for evil. it just feels that much more iffy when you know your research may suddenly become classified and will never see the light of day outside of military use.

i could've been cool building vision for a vacuum cleaner robot or something mundane. eh, but i dropped out and all that.

sorry for long story,
m.

Most of the computer vision stuff at my school is funded by the med school. They want computer vision to identify cancer in lung xrays and stuff. Lots of eye tracking research too for the BCS department.

hey! now that would acceptable! i could get behind that.
m.