bob wrote:diep wrote:bob wrote:diep wrote:hi,
My chessprogram runs on a box that uses a switch that doesn't have a 'power off' button.
This mellanox switch however, you can see from my facebook pictures there (my entire facebook is open to the entire world to browse), it produces 75 decibel or so and i sit a few meters away from it.
Now i got rid of the fans from the switch and replaced it by cardboard and big outside fans. Where the psu can give power to 1 fan that cools the psu (hopefully it'll go ok as the airpressure of the fans i use isn't that high), the other 3 fans from the case do not have this luxury solution. The easiest solution is use an external adapter.
This external adapter could break at which point a disaster might happen to the switch. From 1 of the motherboards of the box i can easily monitor the RPM of the external fans. If one of them drops to near zero, which would happen for example if the adapter breaks, then i want to give the controlling proces there the order to cut the power of the switch.
Writing the software i'll manage, but what device to use that i can control by software or maybe even a simpler solution is there.
Note that i do realize this is not a 100% solution, yet the odds that one day the 12 volt adapter breaks is there, so i want to catch THAT possibility.
I feel some cheapskate solution probably exists. Any thoughts on how to solve this i welcome.
This linux, windows, or some other custom o/s???
In linux it is pretty easy to write a shell script to access the lm_sensors info, and if a fan's rpm drops to some chosen value, shut the system down quickly...
Right lm_sensors it is. Problem is not the software. Problem is how to interrupt power of the switch. I need a device that can react onto the computer.
Now i can of course buy a legorobot that can do it using robot-OS - maybe there is something cheaper.
The switch doesn't have a 'turn off' button.
So i need to physical build something that can be steered by computer. Any thoughts on that?
OK, the switch is a separate box, completely. And you want to be able to turn it off from the computer?
Correct!
For a $100 switch you don't care of course but for a superior mellanox infiniband switch...
If I were going to try to build such a thing, I might "go simple here." Take one pin + ground from the parallel output port, and gate that to a simple darlington-pair power transistor that is used to ramp up the current to a level that will trip a relay. Then simply wire the spdt relay in series with the hot wire on the a/c power, and you are set. Output a 1, the relay closes, and power comes on. Output a 0, the relay opens, power goes away.
If you want it to "default on" you could toss in a "not" gate on the signal coming from the parallel port output, as it almost certainly powers on with an output of all zero bits. Notting that would give you a 1 which would start things off with the switch power enabled...
I have no idea what you're talking about Bob, but from what i understand you want to start getting from the mothballs some solder and other stuff and make your own electronics?
No easy to use ready to use solution from a store?
I see a relays is $23, can i plug into that something else?
The switch is a 36 ports managed QDR switch from Mellanox.
They upgraded a supercomputer to FDR somewhere and i got the QDR switch from it.
now QDR is a great network to use of course. the new FDR for your info (sponsor talk) is of course the superior new network. I'm gonna write a proggie to measure the hashtable latency you can expect and will ask mellanox to also run it at a FDR setup of them.
expectation is FDR is 2x faster in latency (it's quite a tad faster in bandwidth as well, but that's no big deal for us).
I heard timings of 0.85 us for a RDMA (remote dma read from other motherboards RAM). QDR is not even remotely in the same league.
Mellanox has FDR, Qlogic doesn't have QDR and won't have that any soon, also their statements reflect that. They're busy improving latency of QDR a tad, but that will never be even in the same league of the FDR latencies.
Around a 0.5 us worse.
Now all that is single core latencies of course, i'll make a test for all cores to all cores, which is what we want to know of course; expect way worse latencies there than the good weather ones from all manufacturers we saw so far; probably FDR is gonna own everything there. Yet i am very happy wth this QDR switch of course. Many supercomputers aren't even in the same league of those bandwidths
it's 2 x 40 gigabit both sides.
there is no way the mainboards i got can handle that bandwidth
On paper it can do 8GB /s (pci-e 2.0) but i really doubt it will do it