An oldie ... drastical change

General Chia Discussion
1

I have several systems plotting as farming but one oldie … my oldest and just well to do somehting …

So the thing is still plotting even as it’s 8h/plot , hey old so don’t say wtf or whatever …does other things also as oldie eg a low theta node, which makes like $5-10 a month so it pays it’s electricity back

yea granted to old maybe but hey it’s an extra plot simply as for doing nothing, now I finall updated on that super old system the chia gui and result is plotting changed to 14~16h no different settings still 8 cores, xeon same 64 GB mem

So wondering what could have changed that it went from 8h to well 14-16h/plot

I didn’t mind the 8h plot was well for a system that is barely usably anyway present day as workstation

2

Which Xeon do you have on that box?

On my Xeon e5-2695 v2 (12 physical cores), I am getting ~40 min plots, although with 128 GB RAM. If you are on Windows, take a look at RAM caching software (e.g., PrimoCache - I used it before upgrading RAM), and set it to cache your t2 folder (in MM). I guess, you can do it with Chia plotter as well to get a speed boost.

Also, sorry I don’t know anything about Chia plotter, as I am CLI MM only.

3

oldie as said a W3520 can’t compare, who ever would plot 1/8h lol except me

4

OK, so that is just 4 physical cores. I guess, maybe you are also using HD as your tmp folder.

Still, check out those caching utils, as you have plenty of spare RAM to make a difference. Although, you would need to split your tmp folder into two, as the smaller one is where the most hits are, but the bigger one holds big files that are not used that much, but rather choke the cache.

5

meh not gonna work on it to make it better, simply wonderd why from 8h it went to 14h+ and that happened only after updating from like gosh I frgot 1.2.6 or something to the .10

6

Dang dude!? You running Windows on this machine? If so, you should really think about ditching Win for Linux. I have a very similar setup with 2x e5-2690 v2 and a 110GB t2 and I’m putting out like 26 min plots. Just saying. Idk how in need you are of plots fast, but you can cut your times down considerably if you want.

7

:slight_smile: Actually, not. Those tests with PrimoCache I did on a single processor fairly new box. When I got that dual processor box, Win just couldn’t handle those two processors, so day one I switched to Linux.

Is that 26 mins per processor or per box?

I tried to run just one MM instance, but it was rather slow. Running two MM instances, and each one does 40 min plots, so overall that is 20 min plots per box.

8

Ah! Didn’t realize you were running 2 instances. I was speaking of only one instance. I need to fire mine up with 2 MM instances. I have enough RAM. I wonder what my times would look like!? I’m only running at around 75% or 80% CPU usage max with one instance. I’m gonna give this a shot. How do you do it? Just start the second instance when phase 1 finishes?

9

I think that the reason for that CPU usage is slow RAM, and potentially how RAM / SSDs are being accessed (I am getting about 80-90% with 1886 MHz). In order to really fine tune it, you would need to experiment with NUMA settings (numactl), what I didn’t do (discovered it when I was already done with my plots). You may want to check your mb block diagram.

Actually, I would also check whether you are not thermally choking your CPUs. Initially, I was, as the original heatsinks really sucked, so I switched to a single fan water AIO per processor (temps dropped by more than 10-20 deg, and it is really quiet).

I delay the second instance by about half of the total time (just give it 20-25 mins delay). I am also using two NVMes, each in a PCIe slot that is controlled by respective processor. However, without NUMA, it is really a crapshoot game, as we really don’t know which RAM modules are being used for a given RAM drive, and neither which MM instance runs on which processor, actually even if all logical cores that such instance is using are sitting on just one processor.

Still, that is about 70 plots / day for around $1k. Although, as with all older processors, the downside is power consumption - mine draws about 450W when plotting, so I started to build my own coal plant in my backyard :wink:

10

I’m pretty sure my ram is running at 1866. I don’t think my CPUs are being thermally throttled. They never rise above 60 C. I have a Dynatron R5 sitting on top of each one along with the case fans. Should be getting adequate air. I too have 2 nvme’s available on separate PCIe adapters. I feel like I have a pretty good setup to try this with. I guess all I have now is to figure out how to use numactl… Any good sources of information handy?

11

Two PCIe slots is not the same as a given processor controlled slots. Just check your mb manual.

Just try to run two in parallel, and see whether you can get better timing.

As far as numactl, just man page. I didn’t really spend any time on it yet.

12

Yea. Half the slots are controlled by 1 cpu and half the other. I have a Supermicro X9dri-ln4f+.

13

An update on my experiment in the event that you care lol. So, I got setup with numa and splitting my RAM. Began plotting with 2 instances of madmax about 24 hours ago. I don’t know the exact numbers, but I’m probably yielding another 25% of plots per day! As of right now, I’m pumping out 2 plots every 39 minutes! Old setup with single madmax instance was doing 1 plot per 26 or 27 minutes roughly! Thanks for the suggestion.

2 Likes
14

Great update! Next time I fire my box, I will start with NUMA setup first.

Also, did your CPU % got better?

15

Yeah. It got a little higher usage. I still couldn’t get it to peg at 100% in glances or any other system monitor, but higher yes. Probably 92% peak usage. Either way, it’s pumping out a good bit of plots more so I am happy! Thanks for the suggestion.