Old Intel Xeon servers for plotting: how old is too old?

Not sure if the server version is inherently faster, but you could certainly avoid running the desktop to save some resources without reinstalling, simplest way:

# systemctl set-default multi-user.target

and reboot, then if you want to change it back

# systemctl set-default graphical.target

and reboot

1 Like

Here is my cheap plotter. By far the best $ per plot performance I have spent.

The server is approximately 7 years old. Here are the specs, cost and current performance.

Supermicro 1U X9DRI Server (X9DRI-LN4F+ Motherboard)
Dual E5-2670 V2 2.5GHz Processors (20 Cores)
128GB RAM (12800R ECC Registered Memory)
DVD Drive
560w Power Supply
4 Gigabit Ethernet Ports
Crucial 500GB SATA (boot drive)
Samsung 980 Pro 1TB (plotting drive)
NVME PCIe Adapter (RIITOP w/Heat Sink)

Total cost: $814.35
Current Production: 31 Plots per day

Currently waiting on another 32GB of memory so I can attempt plotting with a RAMDrive. Should arrive in the next week. That is an additional $122.95.

Thanks for the great info.! Once I move my farmer to a different machine I will give it a try and see if there is any improvement.

Doubtful that desktop or server makes a difference. What processor(s) do you have? And what amount of ram? Faster ram would probably help - but depends on your processor. I’m getting 25 minutes on Dual E5-2690v2 on crappy DDR3 1333 Mhz ram. I’m pretty sure that system is close to the limit, but with a newer generation processor and DDR4 ram, definitely faster is possible.

You can switch off the gui though, even with the desktop. If I remember correctly, you should be able to hit like ctrl-alt-F2… Or maybe control-alt-backspace. I don’t use desktop so not sure on that, that might give you a few more seconds too…

I’d be careful about adding just 32GB of ram. Right now you probably have, 8 slots populated? That means you’re in one of the optimal configs of running 8 channels (4 per CPU). Unless you add another 8 sticks of memory, you might go from a great memory config to a crappy one.

With 128GB of ram, you should be able to run a 110GB ramdisk already for temp2. I’ve heard people doing that just fine.

My current 128GB of RAM is all 8GB sticks. I am adding 4 more sticks. Balanced across the 4 banks. Should be good.

You can give it a try but I’m not sure that will work properly, because it’s 4 banks per CPU, and using 20 slots sounds unbalanced. 8, 16, 24 with dual CPU or 4, 8, or 12 when using only 1 CPU.

I’m running dual E5-2680v3. Currently have 128GB of Ram and going to double check speed tonight.

With those processors, ram speed is probably the best guess for going faster. Are you using discard mount option, and/or using fstrim? I suspect mount without discard and manual fstrim between runs works best.

After I add the memory, I’ll make sure I run a plot the same way I did before to test that the speed is the same before making any changes.

I just got the chance to check ram speed and as I suspected it is 2133. So I guess my plot times seem to be in line with my specs. I edited systemctl to have fstrim run hourly on all my drives.

I just bought another Dell T7910 on FB marketplace this week for pretty cheap ($500 and they threw in dual monitors, PCIE NVME card, and a bunch of other things) . It has super basic specs. but soon I’m going to upgrade the processors, ram, and throw in an NVME, and then run it side by side with my other machine.

Hey, when I’m wrong, I’ll admit it. @dynafire, you were right on the money. I added the additional 32GB (4 more 8GB sticks) and my speeds got worse. Even a direct comparison using the same command as before, it was much, much worse. Needless to say, RAMDisk with this configuration was useless. So I had to remove the memory I added. The only way I could make it work would be to get another 32GB. (I can’t just run 110 RAMDisk the way I have it setup because I’m on windows with WSL Ubuntu. Just not enough free RAM.

But as a benefit, by accident, I moved my NVME to another bus and I gained speed. So to try something else, I added another NVME and split temp1 and temp2 between them and now this old server is my fastest plotter by about 100 seconds. Even without RAMDisk.

haha I appreciate hearing the update… And glad you were able to find some improvement in another way!

So with WSL, I can understand your hesitation to use a ramdisk… But, if you are willing to experiment you might want to give it a try anyways.

Here’s what I’m thinking. I run WSL on my laptop, though don’t use it for plotting. Right now, my WSL2 debian instance reports 20GB of available ram, and 29GB of swap. The swap is a file in windows, which happens to be an NVME drive. So my swap should run at approximately nvme speed.

So if you create a 110GB tmpfs in WSL, and you run out of ram, it’ll just overflow into your swap - and if you ensure your swap file is on NVME, then it should probably work just fine. You might need to double check where windows is putting your swapfile, and you might actually need to lower the amount of ram the WSL instance has (so that windows itself doesn’t get swapped out), but it seems plausible to me that it would work and actually be pretty fast.

1 Like

I get away with dedicating 110gb ram for Tmp2. Temp1 drive is one is a nvme 500gb…
the setup script on madmax says 110 go os what’s needed. I give another 10gb I give to the os itself than just use 512 buckets (ish) to keep the ram usage down. 24 threads.

Have you tried 256 buckets vs 512 buckets to see the time difference? I tried 512 buckets once but it crashed mid plot and I haven’t tinkered with it since. I used the same settings I had with 256 buckets (24 cores/threads) and I think if I lowered it then it would probably work, but then again I am not sure if my times would improve. But overall I’m pretty happy with my 25~ min range so I’ve left things alone for now.

Havnt tried 256. Will now tho. I only have 124 gb ram available.
I run on my dell t710 with x5690’s
1container harvester
1 container mad max
And full node GUI virtual machine.
All Ubuntu.
I noticed my response times to challenges are a lot better and I get more plots passing filter with a harvester handling the mounts. But once the plotting power is not needed. I plan on making a container harvester on each of the drives I have… 70 tb worth. So each harvester will be in charge of roughly 40 plots. Kind of like how raid is faster… is my thinking… multiple answers from different sources at the same time…. Instead of one full node handling all the drives individually.
It’s a jbod setup with a few servers plotting.

But I’m just a dreamer haha who knows the keys to this chia madness. I try somthing new daily. Probably costed me days of time. But I’m just after that sweet spot. Where the chia just flows in. … I was only passing filter once or twice every few minutes. Now around 4-8 with my 1 harvester.

30 ish min plot times. Can’t complain.
Considering running 2 vms in parallel lol with mad max…