Bladebit Cuda v3.1.0-beta1 - Extremely Slow Transfer from Memory to NVME

Support
1

I plotted with GH a while back and I’m moving back to official c7 plots using bladebit cuda. My system is a Z420 with 256GB of ram, 2650v2, and 2xNVME 500GB in raid 0 mdadm (ext4) for my temp driver and os is on another ssd. I have 3x16TB drives that I’m using plow to move the plots from temp.

This same issues happens on v3.0.0 as well.

The problem is that when a plot is finished its taking 500+ seconds to move from RAM to my temp drive. The first plot will take only 30 seconds. Then 80 seconds. Then 500+ seconds. And will typically just sit around 500 seconds for every plot after that. When the plot finally is moved to the temp drive it’s transferred to one of the 16TB drives in ~5 minutes.

I have tried both ext4 and btrfs for the temp drive and its the same issue.

2

So try this
Releases · Chia-Network/bladebit (github.com)

3

I will give it a try and see.

4

Same issue persists with rc2. Plot creates in about 5m, but the transfer to temp drive is slow. I’m watching the temp drive and the transfer rate will put it around 600 seconds to complete.

With iotop I can see it’s about 100M/s transfer to the temp drive.

5

Ssd slowing down over time sounds like a trim/discard issue.
You could also try running with a single nvme to rule out any issue with the raid setup

1 Like
6

it is a known bug. Suppose to be fixed in the testing version. There is currently kludge to have RAMdisk, and direct copy to JBOD. Storing on SSD/NVMe takes 80s on my rig. RAMdisk > JBOD <15s