Raid across 200 drives

[Lowlandmink543]

I have had the thought of making a large raid array. I was thinking of buying 200 usb-c thumb drives, then making 1 huge raid array. The picture down below shows an idea of what I was thinking. Does this make sense and if so then will this work theoretically?

200 drives total.pdf

[RGProductions]

1 minute ago, Lowlandmink543 said:

I have had the thought of making a large raid array. I was thinking of buying 200 usb-c thumb drives, then making 1 huge raid array. The picture down below shows an idea of what I was thinking. Does this make sense and if so then will this work theoretically?

200 drives total.pdf

theoretically I think it could technically, work, but the real question is why? Also, you would have to find a way to keep all of that cool.

[Electronics Wizardy]

that works, but is extremly messy, so I wouldn't do i like that.

 

With that many drives, id look into a distrubted file system over the network like ceph, storage spaces direct, vsan, gluster, lustre, moosefs and others.

 

2 minutes ago, Lowlandmink543 said:

have had the thought of making a large raid array. I was thinking of buying 200 usb-c thumb drives, then making 1 huge raid array.

Those drives are very slow, just get one good ssd and its much faster.

 

If you want all those drives on one system id just use zfs and make a ton of raidz2 or raid z3. Don't nest parity raid.

[Lowlandmink543]

From what I understand, this would be possible but is not convenient. How would I apply that many levels of raid to the drives?

[Electronics Wizardy]

5 hours ago, Lowlandmink543 said:

From what I understand, this would be possible but is not convenient. How would I apply that many levels of raid to the drives?

well zfs would be much better here, make one pool of lots of raidz2 or raidz3

 

but if you want to do this on linux the bad and hackly way is to use md to make the arrays, then make a md volume of the md volumes, and so on.

 

But don't do that with nested md, thats a bad idea.

[aspoelst]

Why?

[dalekphalm]

On 6/26/2019 at 5:26 PM, Lowlandmink543 said:

From what I understand, this would be possible but is not convenient. How would I apply that many levels of raid to the drives?

What are you trying to accomplish, precisely? That will help us answer you better. 

 

Many hardware RAID controllers have a physical drive count limit (even with expanders), so you may hit that point first. 

 

Also many won’t let you nest that many. 

 

But that could work under certain circumstances. 

[PianoPlayer88Key]

The idea of 200-drive raid gets me thinking ...

This would be a bit crazy and I wouldn't try it myself (but maybe with Linus you never know), but...

 

Theoretically, how many MFM / RLL (ST-506 / ST-412), SCSI,  IDE / PATA, SATA or SAS hard drives, or 5.25" / 3.5" (native, not USB) floppy drives, or other older / slow non-flash storage devices, in Raid 0 or whatever, might be needed to match the random IOPS or MBPS of a single high-end NVMe SSD (like 970 Pro), Optane SSD (DC-P4800X?), quad-channel DDR4-4500 RAM, GDDR6 / HBM2, or level 1 CPU cache on a flagship Skylake Xeon or Rome Epyc system?

 

(Not counting latency, which I'm sure regardless of how many drives you stripe, would be dominated by the NVMe / RAM / etc., But just raw throughput but random, not sequential.)

 

One way I was thinking to test it, without having to build an array with 9001 drives or whatever, would be to benchmark 1 drive, then 2 in Raid 0, then 3, 4, on up to 8 or something.  Observe the performance between them, and see what pattern emerges, then extrapolate up.  If there comes a point of no further gains (for example having 500 drives is 0% faster than having 25) where does that occur?