This is how cache works:
it is a relatively small buffer to ensure very fast writes. However it can only do this for a very short amount of time. Under normal conditions this is very useful, but not for plotting.
A single plot writes about 1.3 TB of data, compare that to the cache of 12GB and you know it will fill up quickly. More importantly you never really stop plotting so you never give the cache time to recover.
In this picture you can clearly see the effect of cache on the write speed over time:
The SLC cache is exactly how they achieve faster write speeds. It’s considerably faster than TLC.
You are also ignoring how Plotman works. It actually does not just sit there and write one large file to the disk the entire time. It actually computes hashes then writes hundreds of smaller temp files, the number of which is a function of the bucket size setting. It then pauses to read (scan) values from those tables, then it begins writing again. The largest sustained write occurs at the end when the last batch of temp tables are stitched together into a plot file.
I’m not going to debate this further… You’ve got your opinion and I have mine. We’ll have to agree to disagree.
yes for brief bursts, that’s why it’s called cache…it caches stuff. Cache is a buffer nothing more. It does not never ever actually increase performance of the underlying hardware.
yes and this completely overwhelms any ssd cache you might have.
Nope, most of the writes are in phase 1, with brief pauses or not the metrics are all here on the forum, people have measured the shit out of this.
It’s not a matter of opinion but a simple matter of metrics and the laws of physics.
You have to write 1.3 TB of data into the cells of your ssd, this takes an x amount of time as per the specifications of those cells. Cache doesn’t change this.
SSD’s with low sustained write speeds, suck at plotting
Again, you don’t have to believe me, this is a well established case.
I’m sorry to go on about this but looking at ssd write speeds that are only achieved by the cache speed was what got a lot of people in trouble 10 months ago.
SN850 is good for plotting, but that has very little to nothing to do with any cache it might have
Not in my experience, but I’m using RAM for tmp storage. I see heaviest writes in the later phases.
Hence the 300GB tier of SLC storage in the SN850, which is exactly what I referenced. That is why that device is ideal for plotting one or two concurrent plots but falls down in performance after that.
@Voodoo what both you and @Shackrat say are correct, I dont even think you guys are arguing on the same point.
I have used madmax with tmp1 and 2 on SSD, or tmp1 on SSD and tmp2 on ram and the behavior is that bunch of smaller files are happening and once a phase/section is complete it writes a chunk of sequential (usually 20-24GB) during end of phase, then combines them near the end.
if one is to use tmp1 AND tmp2 both on SSD, then obviously the cache would matter that much more hence what @Voodoo is stressing on, which is also why early stage people have problem, because nobody used ram as tmp that early until madmax came along.
however with madmax allowing tmp2 in ram, the SLC size on these QLC/TLC SSD become much less of an issue, because the cache actually has time to dump files onto the QLC/TLC portion while rest of the plot is doing other things thats only writing to tmp2 (assuming plotting is not fast enough to overwhelm SLC cache). this is what @Shackrat was talking about.
both are right in their respective use case. in general I’d still like to avoid SSDs that has low sustain write because feels like im buying something with a masked performance, tho for daily usage on OS it is totally fine.
Yes you make a good point, I have to correct myself a bit here.
I guess since the advent of mm, we are only doing 1 plot at a time and cache might have some use, especially when you are using part Ram like you said.
But, whatever you do, you still need to get a ssd with good sustained performance. Any cache benefit is a small bonus at best would be my guess. The case where you are doing 75% plotting in Ramdisk, the ssd is much less important anyway. But if you use only ssd for plotting, you need to select carefully.
In any case I would go with confirmed good ssd’s, whatever the specs. The benchmark for plotting is…plotting!
yea with the way industry is moving towards with SLC paired with QLC or TLC or the upcoming PLC from intel the sustain performance is getting worse and worse due to flash performance gets worse as it gets more divided.
to have high sustain performance with shittier cells like QLC/PLC, one would need to have them in multiple “channels” which would mean a more expensive SSD controller, more complex printing on the PCB which is more cost (not to mention higher overhead latency hence the use of SLC to mask the extremely poor latency, something we’d never find out unless we test QLC/PLC directly).
some of the 4-8TB M.2 SSDs from companies like Sabrient, already have such a shitty sustain write like 100-200MB/s, not even sure the point to them. Actually, they might actually be good for chia plot check with lower random read latency than hard drive, with their size “16TB” in the future, with only one exception, price.
Sure but it defeats the whole point of not trashing an SSD (albeit far less than plotting directly to an SSD) when plotting in memory.
I have 384GB in combination with madmax and have 26 min plot times (+10 minutes plot transfer time to HDD) which is fast enough as I extend my farm disk by disk, no SSD is being murdered while doing this .
True but depends what type of SSD you are using, this is not a heavy workload for an enterprise SSD. Also this setup is not primarily used for CHIA, once plotting is complete it is used for its primary purpose.It is just fun to see how fast we can go. When I started in May I used the GUI to plot and it took 4-6 hours on a workstation, then used MM on same workstation and got it to 29 mins and now 5 mins on an enterprise server… I can actually get it lower if I use an Optane drive… write to disk takes 6 seconds
wow, that’s some fast HW you got there
Would like to do this also, but the investment for this type of HW makes no sense for the sole purpose of farming Chia. Anyway I would like to to see if we can get less than one minute plot times in the near future (Epyc ZEN4 or latest Xeon HW)
So one would need an intermediary SSD to move the plots, defeating the whole point of plotting in RAM.
or can just do 8x SAS hard drive in raid 0, which is cheap like $200 and it’ll give enough write speed faster than plot write, then pair that up with a 10g network transfer as well as destination of 3 hard drives in raid 0 call it a day.
The larger capacity NVMe drive will probably be faster. How much faster? I do not know.
The reason for the faster speed of the higher capacity drive is due to the drive having more storage cells / chips / NANDs (whatever it is called). So when data is being written to the drive, it can be sent to more places on the drive, simultaneously (akin to a RAID 0 built into the drive). Same thing for reading data from the drive.