Slashdot Mirror
Domain: myrinet.com
Stories and comments across the archive that link to myrinet.com.
Comments · 7
-
Re:Low-cost options?Define low cost? Myrinet with less than 10 microsecond latency is normally considered to be the least expensive option. You can check their price lists, but an 8 port solution (with 8 HBA's) will set you back over $8k, not including the fiber.
For some people, that's cheap. If not, sorry.
-
Re:Low-cost options?Define low cost? Myrinet with less than 10 microsecond latency is normally considered to be the least expensive option. You can check their price lists, but an 8 port solution (with 8 HBA's) will set you back over $8k, not including the fiber.
For some people, that's cheap. If not, sorry.
-
Re:Authentification
Seti@Home already blows away all other supercomputers on the planet
Where did you get this idea? I'd honestly like to know. It's very misinformed.
There are many problems for which Seti@Home distributed-style computing would be worthless. One of the major selling points of real supercomputers are the interconnects between nodes. Big problems require fast interconnects so that nodes don't have to wait for data from other nodes working on other parts of the same problem. The Internet is way too slow for problems that have massive amounts of data that need to be passed between nodes -- such as modeling earthquakes or the earth's magnetic field.
-
Re:You misstated the latency issueTo cite the article: The IEEE 1394 bus has a minimum latency of a few hundred microseconds and a worst-case delay of a few milliseconds.
Normal latency for SCI (Dolphin delivers SCI) is about 1.5 microseconds; for Myrinet, about 8 microseconds. We are talking about latencies that are three order of magnitude (a thousand times) smaller.
FireWire is not enabling for the same range of applications as a full-blown scientific cluster interconnect is. However, few applications need s a fullblown interconnect - most clustering apps I've built could have been tied together by the bandwidth of smoke signals (fortunately,they seldom have been - the one case I can remember was when a customer insisted on dictating hardware...)
Eivind.
-
Disks Undesirable; Appropriate Applications
Diskless nodes are much better from the point of view of maintenance and upgrading software on the system. Do it once, on one node (the "master" node with the disk), and all nodes are updated.
Caveat: Using the network for temporary storage hurts performance and doesn't scale if you have a lot of temporary data.
High Performance Computing Hardware
If your parallel process has a lot of communication, you would be using Myrinet, and/or getting a CSPI cluster running Linux, because ethernet is too slow. (You would avoid Mercury and RACEway because they are avoiding Linux.)
Appropriate Parallel Applications:
In short, this system would be good for embarassingly parallel tasks like a brute force search through key space for breaking encryption, or encoding mp3s. In such tasks, each processor only has to communicate at the beginning and ending of the task, and no processor has to communicate with any other processor during the task.
Inappropriate Parallel Applications
This system would not be so good for STAP in RADAR or SONAR.
Kenneth J. Hendrickson
-
Re:Problem with current programs
See also: MyriNet
-
Re:What kind of network speed do you need?
How about something like Myrinet? Iffen you gots the $$$
Your Working Boy,