Greetings, I have an amd ryzen threadripper 3990x, 64 cores and 128 threads, x2 SSD Micron 9300 max, 256 gb ddr4, os windows 10 pro, 10 gb ethernet.
I have in parallel, I use a micron in temp1 and a ramdisk and another equal plotter with a micron in temp1 and a ramdisk
I do tests and I do not spend 26 minutes per plot, I would like to do it in parallel, how it is done and what do you suggest about the configuration in madmaxie.
My best results were with the following configuration: Number of Threads: 16 Number of Buckets P1: 2 ^ 8 (256) Number of Buckets P3 + P4: 2 ^ 8 (256) 2 in parallel every 40 minutes.
Why use low thread count if you have a threadripper?
use 32 instead I would say for each plot. (or use -K 2 switch to increase the threads after phase 1 )
In any case 26 minutes is not very good time for this system I think
One thing is that ramdisk in Windows is almost always much slower than Linux, you might want to try Ubuntu.
Itâs my understanding that the -K flag only affects the number of threads for Phase 2 (and does so as a multiplier).
I agree with Voodoo. Youâve got 128 threads available, is your CPU utilization maxed out running two 16-thread jobs?
Have you tried different bucket values? Increasing the number of buckets to 512 (or even beyond if your rig likes it) will reduce the amount of RAM required for each thread by half, therefore you can run more threads on what RAM you have left over after allocating for your ramdisk.
Did you increase your threads when you ran 512 buckets? Because thatâs what it allows you to do â when youâre limited on system RAM from a ramdisk, increasing the buckets makes plotting use less RAM per thread, allowing you to run more threads.
Just cause it seemed to me that your main limitation is RAM, not threads. Youâve got plenty of unused threads but you donât have enough system RAM for all of them + OS overhead. More buckets = more threads with RAM being equal.
Beyond that I have found much, much better parallelism between MM instances (on my ThreadRipper) modifying and fixing process priority & core affinity settings. That is, start (either two or four) Powershell processes (depending on which above is used), then go to details in task mgr and give âhighâ priority to all these PS instances.
Next allocate the correct number of cores with affinity (either 32 different cores to each of 2 instances or 16 different cores to each of 4 instances). Last, of course, setup each MM PS command string in its respective PS instance before (pretty much) starting them all off consecutively. Or you might stagger them by say letting (1 or 2 depending) finish Phase 1 before starting the other 1 or 2.
I realize you will not be utilizing your RAM to the fullest or even close, but it will show you how much, if any, RAMdisk is actually helping. Furthermore, you will be able to run 2 or 4 instances without regard to running out of memory, as you will have far more than necessary without fiddling around to get it âjust rightâ so as not to use too much, such as when RAM disk takes up most of what you have.
Lastly, saving off to disk will be faster if you have 2 or 4 to write to individually, but in the end it doesnât matter much, as itâs such a minor load on your SSDs and/or your CPUs during that transfer.
6.4TB 9300 Max has a 3 Drive Writes per day rating, so 9.6TB/day and 37.3 PB endurance or 37,300 TB. Thatâs very generous amongst SSDs, and @1.3TB/plot more than a few plots, like over 28,000 plots! And you have two, and endurance is generally way understated. Donât think you have much to worry about in this lifetime
Iâm doing over 6 tb in one day on exactly 1/4 your cores and two MM instances, so you potentially have plenty of room to improve. Maybe using the RAMdisk isnât so great as you imagine, just perhaps?? I only use two lowly consumer 1tb SSDs & 1 2TB SSD all PCI-e 4.0, no RAMdisk at all. Nothing hardware compared to your setup.
But your SSDs, your workstation, your choice, no pressure, I was just offering something to try re: âany recommendation welcome, to improveâ. My thought is experimenting and trying all alternatives is what makes this all so much fun. Enjoy whatever works for you!
Grateful for your guidance, believe me Iâll try, if I do that configuration d x4 micron 9300, it would be two in phase 1 and the other 2 in the other phase, right?
That will work, but doubtful youâd have problems even with all in phase together with those fine SSDs. And note: the â-G trueâ parameter switches SSDs between -t and -2 after each plot to even wear. Just try one way and then the other to see if any difference.
No, youâve used all your cores (64) in the first 4 instances. You want to keep -u to total only cores available. Youâll see that i.e., a 16 thread instances ( -u 16) will use up to 32 threads in phase 2, and even some in other phases beyond 16 occasionally.
Just keep in mind over specâing threads will slow everything down, to the determent of plot speed.
You could replace one of your 9300s if the 4800x is big enough to handle 4 instances going simultaneously. or do some on the 4800x and some on the 9300.
. My thought is experimenting and trying all alternatives is what makes this all so much fun. Enjoy whatever works for you!