K35 plotting tweaks?

First of all I want to thank everyone here with all their thoughts and the tweaks guide up top has been helpful in trying to get this running as smooth as possible. Before anyone asks why K35, I think it’s the right solution for me for the long term. I’m very much in the mindset to plot as few times as possible, even if it takes longer initially.

My specs:
CPU: Ryzen 7 5700G

RAM: flexible, 32GB-128GB DDR4 3200

SSD: flexible. Firecuda 530, Corsair Force MP600 or Inland Premium 1TB drives in a 3x1TB RAID0. I would rather keep all drives the same in RAID0 arrays, but if someone thinks it might squeeze more performance mixing brands as they are all the same capacity I am open to experiments.

Destination drives: easystore drives via usb3

For my first run I didn’t want to tinker around too much so I left most options in the gui as is. I did allow for 14 threads on the cpu to work, but from what I’ve read you might actually lose performance with ryzen processors over 8 threads with 6 threads appearing to be ideal. Upping the Ram a little bit might help as well.

Right now I am at about 28 hours for the first plot and 73% completed. Too slow? Fast, just about right? SSDs in this run are the corsairs and the SSD utilization rate has been fairly low. I’m not sure what to make of this, suggestions?

It also appears I might have an issue with my usb3 port that has my drive connected. System has been rock solid, but the power on count on my easystore is already at 30. No way this is normal, right? I will shuck it and install it internally if it is.

Second run I will up the RAM and set the thread count to 6 and see what that does. I am currently running the system with 48GB RAM installed for those wondering.

Hi there and hello again.

I know you can’t be talked out of your K35 adventure so I won’t try;-)

What OS are you running and what chia plotter do you use?
Is the 73% progress for the whole plot or just phase 1?

Assuming it’s the standard (chiapos) plotter consider this.

Your CPU is about on par with my i7 for cores/threads and clockspeed so we can compare things.
A K34 with madmax plotter takes little over 4 hours on my plotter, with about 1.7 hours spend in phase 1.
For your K35 that phase 1 would multiply by (very) roughly 24 as a K35 takes twice the calculations of K34 and the original plotter is single core for phase 1. I used 12 threads in madmax, so 2x12 longer phase 1… very roughly as said.
That would mean some 40 hours for phase 1. If the 73% ready is for phase 1 that’s in line with expectations, if it’s 73% ready for all 4 phases it’s very very quick.
Using the standard plotter the trick will be to ‘stagger’ parallel plots, optimizing the usage of all those cores your CPU has with parallel phase 1 calculations. There’s quit a bit of info on that trick from ‘the old days’

On the other hand, it’s pretty early in the morning so my math could be way off… :grin:

1 Like

You’ve been a great resource @xkredr59! As far as the percentage goes, I am just going by what the front facing GUI is showing at 73%. I will have to check the logs to see what’s going on back there. Yes this is the standard chia plotter. This is on Windows.

Is there anything to make sense of the low SSD utilization rate according to windows task manager? Is this just the way it goes? I was fully expecting the drives to be hammered to the max on a consistent basis, but so far wasn’t the case. I would say it hovered around the 25%-30% rate on the task manager.

Do you think it’s possible to stagger with storage considerations? I saw my storage free down to about 800MB free at one point.

A piece of advice for those as dumb as I am to try and attempt this. So far I have to say I think it might be best to go with 2x2TB drives even though the upfront cost will be higher due to the size of the blockchain itself. If you are planning to plot many of these offline then a 3x1TB configuration makes sense. I just try to squeeze the most performance out of machines and remain flexible as I learn.

I’m also interested in possibly setting up RAMDisk to help alleviate some wear on the SSDs themselves, although I am not sure how much benefit that will bring me on that end. Any advice will be appreciated. I am targeting 52GB of RAM on the desktop to keep things as cost effective and flexible as possible.

This will solely be dedicated to plotting so nothing to be considered to get the job done outside of the resources necessary for Windows to run as intended.

If it’s still in phase 1 (I think so), there’s only one cpu core doing the calculations, so the ssd isn’t used to it’s max.
In phases 2 and 3 other cores you assigned will kick in and the ssd will have to work a little harder;-)

Ramdisk is only feasible for K32 plots, for all but very high end systems and even for K32 you’d need 128GB.

As for staggering several K35 plots I overlooked the impact on temp storage needed. One plot alone takes over 2TB of space maximum at the end of phase 2 and declining thereafter to ~1TB after phase 3, so overlaying several staggered plots will quickly grow over your available space. It will be very hard to get the timing right for even 2 plots.

I’ll stay away from pro’s and con’s of K35…

1 Like

Thanks for the response. A quick update I am at about 85% & storage space quickly coming back at 1.40TB free or so.

A few last things. Do you think lowering the thread count from 14 to 6 makes sense? I’ll increase the ram available to it as well to see what that does in terms of timing.

What software do people on here use to monitor TBW and overall drive health? I’ll be interested in seeing what pops up on my end.

On windows perhaps your ssd manufacturer has brand specific software (Samsung has i know).
Otherwise you can use the very good CrystalDiskInfo.

Free download

1 Like

Sounds good, is your ssd-usage up from 30% now?

1 Like

I would try 8 first as a benchmark, the number of physical cores of your CPU

1 Like

Disk usage is around 6%-10%

A little bit from the log. It appears this is compressing tables 4 and 5 currently.
Progress update 0.850

I don’t understand any of this at the moment, but trying to learn quickly. My hobby has always been to build flexible systems that can accomplish tasks as efficiently as possible. I believe I am at about 34-35 hours or so now since starting this plot.

compressing sounds like phase 3 of 4, 7 tables in total to do so you’re getting close :+1:

let us know how it ends…

1 Like

Just in case it is of any use for comparison. I have been switching out some k32 plots for k34.
I have a 1TB SSD for the temp dir and 500GB Ramdisk for temp2 dir. You can use 480GB RAM disk, but occasionally you get caught out and the plot just stops. I use a thread count of 8. I found that using 12 did not make a significant difference. Each plot takes about 3.5 hours plus about 10 minutes to copy it to my NAS.

1 Like

Yes @gtp all info is helpful. Thanks for that.

So I’ve had my eye on the log for half an hour and it went from Bucket 40 to 81.

Uniform sort.
Ram - 28,800GiB
u_sort min - 13 GiB
qs min - 4.105 GiB

Second computational pass time: 3610.193 seconds

CPU (98%)

Wrote 28258848826 entries

Now compressing tables 5 & 6. Percentage is now at 92%

I had to step away for a bit, but last checked was at 98% about 2.5 hours ago.

Final build time was at around 41.5 hours or so.

Realistically how much time can I shave off? File appears to be final at 883.793 GiB, but the GUI is still showing “plotting” at 98%. Final file is a .tmp file?

File has already been transferred to the external hdd automatically.

Safe to shut down chia and restart? Does the temp file auto delete or do we have to manually delete that file off the SSD once completed?

Those running a node on the same machine and attempting to do something this dumb. Go with 4TB+, this is cutting it way too close without doing some sort of Ramdisk solution that I’m not sure how to implement.

So it looks like the file DOES automagically delete after transferring resulting in a “PLOT” file on the HDD. Does it really take 8 hours to transfer a finished plot file to an HDD? That feels absurdly long…I’m starting to question if this external I have is a dud.

Now to continue tweaking…can I get this under 40 hours is the real question now.

Congratulations, you’ve made your K35 :+1:

To possibly be a little bit faster you can play with buffer, threads, buckets and stripes parameters.
Difficult to predict which tweaks will lead to what result. It all depends on cpu clockspeed, number of threads, ram usage and speed as well as temp speed.
For a bigger improvement you’d need a multi-core phase 1 plotter, no such thing available currently.
Or a higher (over)clocked CPU, faster memory and temp storage (4TB Optane would be great).

I thought I wasn’t gonna bring this up again but for your long term chia plans, please reconsider plotting K34’s with the inbuild madmax plotter; expect around 4 -5 hours so 8-10 for a K35 equivalent plotted.
You’d still be way ahead of the crowd and far beyond the longevity of your hard drives before K34’s get out of fashion (if they ever do).
I won’t bring it up again.
I think…

I am considering it as I continue my calculations to spread across all my systems. K34s are not completely out of the question.

I will look into seeing what a 3TB Storage/52GB RAM combination can produce in terms of parallel plotting for those tomorrow.

Good afternoon.
Sorry for the probably stupid question, but what size of the SSD disk is needed for the K35 if all phases are done on it?
Perhaps there is a table of the required SSD for K33-35?

Thank you in advance.

https://github.com/Chia-Network/chia-blockchain/wiki/k-sizes :grinning: