Hi there, has anyone been successful to plot to ram on a 512GB machine with the bladebit cuda alpha?
I want to remove the need for a SSD and then simply offload the plots from RAM to the permanent storage…
Howe when I create a tmpfs on Ubuntu, size 200GB, bladebit throws an error…
Anyone? Cheers!
Someone said before, ramdisk does not support directio
I did a bit of googling and found GitHub - pkoutoupis/rapiddisk: An Advanced Linux RAM Drive and Caching kernel modules. Dynamically allocate RAM as block devices. Use them as stand alone drives or even map them as caching nodes to slower local disk drives. Access those volumes locally or export them across an NVMe Target network. Manage it all from a web API.
But I have no idea about that.
Perhaps a wild idea…
Make the ramdisk
Create a 100G file in it
Loopback mount (or nbd or nfs or ssh) the file
Bladebit uses direct I/O for the copy, and this is not supported on ramdisks.
Maybe there is a workaround or some way to make a ramdisk that can handle it, but by default it’s not possible.
The OG Bladebit can create plots all in RAM if you have 512gb and offload straight to HDD. The new BladeBit CUDA only needs 256gb of RAM to avoid all use of SSD temp space and can offload straight to HDD as well.
Yes indeed, that’s how I use it. I wanted the “offloading” to be to the RAM disk, given I’m only using 256 of my 512 for the GPU plotting.I understand from Voodoo above that it’s not possible due to the “direct I/O” aspect… Shame, would have been nice to do that but I’ll just put a few HDDs in a RAID0 and hope the write speeds are decent…
Thanks folks!
Let me try this, it looks like that little tool does exactly what I need… maybe it can “trick” bladebit into thinking it’s a real drive… thanks !!!
I was able to do that by using the recommendation of creating block device via memory. Please be mindful, if your plotting speed is higher than copying out of ramdisk, you will run out of space and in my experience 512GB was not enough because I needed space for 4 plots in the staging area.
Yet if you want here is how to create memory block device:
modprobe brd rd_nr=1 rd_size=198832178 max_part=1
Command to delete:
rmmod brd
Let me know how it goes
I use Vmware Esxi with an Ubuntu guest with bladebit alpha. Also I use GPU plotting (GTX 1070).
if your plotting speed is higher than copying out of ramdisk, you will run out of space and in my experience 512GB was not enough because I needed space for 4 plots in the staging area.
I was thinking I might be able to create a fast enough RAID with three 18TB Seagate Exos drives that can move the plot out in 3 minutes, the time it takes me to create a plot… I need to see if that can be done, appreciate the concern and thanks also for the code!
Would it be fair to assume at RAID0 with 3 x 7200 rpm SATA drives at an average single disk write speed of 250 MB/s, I could reasonably get to 600 MB/s for the raid config?
Then, the 80GB file would take around 2.5 minutes to copy over, which is less than the three it takes to create it… Am I overlooking anything? Obviously 6Gbit/s SATA connection…
When I run this on Ubuntu, I get:
modprobe: ERROR: could not insert ‘brd’: Invalid argument
Anything I need to do with the kernel to make this work? Can’t seem to find much on this topic on google… Cheers!
Is your raid only for temp drive?
4 x sata is minimum for 3 min plots. DDR3 RAMdisk is simply slower than using SSD. Do not forget, any modern SATA HDD will drop to 120-150MBs at inner tracks. 250MBs is only outside tracks.
with 4 x sata, you can copy directly to drives though. I have seen 80-150s per plot.
SSD will wear out at some point thats why i prefere ram disk.
I have Corsair MP600 2 TB that is almost 5000TBW…officially 3600TBW 
I wouldn’t worry too much 
I know most of them can last out more than the official TBW. But at one point they will break.
Well memory prices also go down, but if you have already have plenty of memory, ramdisk may help.
Did you do sudo? I ran it on ubuntu 22.04 without any issues.
yes, that’s the idea to offload from RAM
if you have DDR4, I am guessing it will be faster than any SSD. DDR3 still works too…I had slightly slower times with DDR3-1333MHz…didn’t test yet with 1867MHz. In my case, it was moving stuff between numa nodes I guess. QPI link is slow I guess.
I have DDR4 2166 modules and 3 fast SATA drives. Will give it a spin, but you are of course right, 4 drives would be better. I just need to come in under 3 minutes… Let’s see…
Yes, but not in 3 minutes like the CUDA version…