Hi everyone!
So… we have a RamDisk on temp2, as we understand, MadMax utilices mostly RAM and so even if we are aware of our small processor, we would like to ask you guys if it would be a good idea to do a RAID0 with our 2 M.2 WD blue of 1tb, do you think it would improve our performance?
Thank you!
I doubt raid would improve things personally.
Using a similar spec cpu, 8350, 8 cores, 8 threads.
I can hit 117 min times with just a single m.2
I just dont think you can expect much better times with that cpu.
Use a RAMDisk. Simply having a lot of RAM doesn’t help. Having it setup as a RAMDisk does. I find that I actually have slower speeds with more or less than 4 threads in the Chia command line. I set my threads at 4, created a 115 Gb RAMDisk, and have a temp NVME and final destination NVME.
Windows 10 using custom PowerShell script I wrote.
MadMax plotter, 1x plot at a time.
- 4 threads.
- RAM A: as temp drive
- NVME B: as destination drive.
i9-9900k at 4.7 GHz (air cooled, Noctua DH-15 dual fans, under 80C)
128 Gb 3600 RAM; 13 for OS, 115 Gb for RAMDisk (A:)
Samsung B: PCIe 983ZET 480 Gb NVME (7.5 PB TBW life)
Plot times: 49.7-50.5 minutes using Mad Max and copying to a network farmer (I use a 970 Evo as a local plotting PC temp swap file for the network copying which takes about 12-15 minutes.)
I get about 29 plots per day on average with this setup. And I can still game while plotting. The only bottle neck is I can’t watch movies when plots are transferring, as they start to buffer on the plotting PC.
RAM A as “temp drive 2” is fine.
I would devide this Samsung 480 Gg into two partitions, one for “temp 1” and the other for plot copying.
What is the advantage? I am guessing no wear on a second nvme? I need a way to plot space used on this HD during a plot (can’t seem to figure that out). My gut is that it’s more than 360 Gb?
Thank you !
You can use 244GB of ramdisk to plot a K32 with madmax. No need to copy before starting next plot.
Use a SSD as destination and you should be good with 256GB on your system.
Thank you, but I have tried with total ramdrive below 270GB and it crashes, which in my experience is caused by running out of space.
It is not important as I have 384GB ram. I have to say that plotting entirely in ram, using ram as T1 and T2 versus using ram for just T2 (110GB) and using AN other drive (in my case I am using raid0 with 2-4 x 15k sas drives as T1) does not yield significantly faster plotting. Not on Ubuntu.
So I wouldn’t advise everyone to go out and buy an extra 256GB ram unless you want to do 2x parallel madmax plots, which I found in windows did achieve more plots/day.
I am also moving plots off to the farmer machine by 1gbps LAN so this is the bottleneck for me. I have reduced to 4-8 threads in MM instead of 28 to save power and match plotting speed to network copying speed.
I do not think it makes all that much difference what drive you use for T1 if you are using ramdrive for T2 especially in linux. In windows it just doesn’t work all that fast and I’ve tried 3-4 of the major ramdrive apps for windows. Some people claim using linux subsystem 1 in windows is reasonably fast, I haven’t tried it.
The bulk of the processing and sorting/shuffling is done on the T2 drive in ram with the completed chunks assembled in the T1 space.
That makes sense.
When I was plotting, my two ordinary machines pouring out 40 plots per day, and I was not hoping for more. That was an all-rounder match (network, copying, stability).
Add:
Now I got the fastest gaming machine I have ever owned.
The problem with WSL is that it does not support tmpfs, so you still need to create a ramdisk in windows first, and then use that in the subsystem. (unless there is another way that I don’t know about)
In WSL, tmpfs just uses a virtual disk file, not RAM
How is the network your bottleneck? Are you copying in parallel or in series? Im running Windows 10. I leave my completed plots, start a new plot and then initiate a copy command across the network.
The copy command is a separate powershell command started at the time of plotting with a 15 minute delay on initiation and a 50 minute offset to match my plotting time.
When a plot is completed it sits for about 15 minutes then I copy it to a temporary destination on the network farmer. Once copy is complete I move the plot to a folder that the chia farming GUI sees on that same drive on the farming machine. The “move” happens almost instantly, whereas the copy takes 15 minutes.
I have no bottlenecks as my plots take 49 minutes each, and my copy + move takes about 15. Even with running two plotters offset, I still don’t see a bottleneck with this arrangement.
Good luck !
Is there a guide for the best plotter configurations and settings for mad max?
I am trying to switch from Windows to Linux after 5 months of plotting so there are a ton of little tweaks I had in Windows that I now have to re-figure out (and remember) to try and get set up again.
For example when using SSD’s for temp1/temp2, what mount options are best? Which filesystem?
has there been any consensus on why mad max plotter suffered performance issues after upgrading the plotter software for NFTs?
For example; https://github.com/madMAx43v3r/chia-plotter/issues/785
a lot of people in there reporting 50% or more performance decreases using the exact same hardware setups just from the NFT updates to mad max; many of them on Windows but some also on Linux. You can clearly see that the CPU and/or SSD temp drives are being heavily under-utilized
another question; why are people still using RAM disks for systems with =<128GB RAM systems instead of a RAM cache + SSD?
If you use a RAM cache in front of a SSD for temp2, you have the freedom to choose less memory dedicated to the cache, giving you more for the OS, and can still reduce SSD wear greatly with as little as ~40GB of RAM cache. So for systems with 64GB-128GB RAM, using a cache + SSD instead of a RAM disk seems like the better option.
Yes, running mad max in a VM on a Windows host is a waste of time. As is using WSL. In pretty much every situation I have seen, using any kind of virtualization on Windows nerfs your disk IO so badly that it gives horrible execution times. The only thing I did not try was to mount the NVMe SSD’s directly inside of the VM/Docker as opposed to shared volume / volume binding them in, maybe that would work better idk.
It’s more work to set up a writeback cache in Linux and I haven’t been able to get it to perform as well as a RAMdisk. tmpfs doesn’t reserve RAM and pages are swap-eligible so it’s a bit more cooperative with the rest of the system.
On Windows I was using PrimoCache, I have not tried yet but my understanding was that in Linux the equivalent is bcache; Bcache - ArchWiki
bcache is one option, but that considerable wiki page only covers using one block device to cache another block device. Before you can even start you first need to create RAM-backed block device using brd or zram. I’ve tried with both; I’ve also tried lvmcache instead of bcache. There’s some potential there, it’s just more work—like I said—and I can’t quite get the same level of performance as tmpfs.
Hi, my plotting time is 30 minutes. I have 4TB of storage on the plotter. Copying across the network to my farmer runs at around 30-60MBps, so even at best network copy speed, I cannot move the plots off the plotter as fast as I can plot them. So I slow it down a little, it generates less heat and uses less power. If I upgraded to something faster than 1gbps between the plotter and farmer maybe, but I’ve plotted 1100 plots like this and am heading to 2000. Its steady.
If I was plotting slower - no problem.
If my network speed was faster - no problem.
That makes cents !
Time for 10gig
Sadly yes, it does only make cents.