Working on an updated used SSD buyer's guide

[Patrick]

Hey STH - I am working on a new buyer's guide for STH. It is going to be an update to: Buyer's Guide to getting a used datacenter SSD inexpensively

So far I have:

• Updated guidelines on what to look for
• How re-deploying enterprise SSDs will revolutionize metal deployments (HDD v SSD impact)
• STH/ DemoEval used enterprise SSDs data set figures for:
  • Reliability data
  • Erasure Procedures
  • Average DWPD
  • Average Power on Hours
  • Interfaces used (SAS, SATA, NVMe)
  • Manufacturers

The data set is now several hundred SSDs versus 78 from last time.

The second point I may pull out to a different article since that can be a big one.

Any thoughts/ suggestions around the above would be greatly appreciated.

 

[j_h_o]

Quick thought: Specific examples of deployment scenarios (specific devices) vs. a set of workloads would be useful/applicable, ideally with some real numbers on what is achievable in terms of IOPS (? latency?) in each case.

 

[Patrick]

Quick thought: Specific examples of deployment scenarios (specific devices) vs. a set of workloads would be useful/applicable, ideally with some real numbers on what is achievable in terms of IOPS (? latency?) in each case.

I was going to address this by focusing later on specific deployment examples. E.g. all flash Ceph.

 

[KioskAdmin]

The original guide I show to clients often.

I'm concerned about quantity of data points. If you don't have 200 drives with a year of use on them is there really a point?

 

[Patrick]

Over 350 drives and a total of over 400 drive years in the tally thus far.

While not a huge data set, it does comprise many different models/ interfaces. And frankly, there is not too much data out there.

 

[RobertFontaine]

Real data is much better than no data.

Used data center ssd's still seem to be more expensive than new retail ssds. I keep wondering whether lots of of little retail ssds in raid are a better value or even potentially better performance for the same dollar than a lesser number of data center ssds?

 

[Patrick]

Real data is much better than no data.

Used data center ssd's still seem to be more expensive than new retail ssds. I keep wondering whether lots of of little retail ssds in raid are a better value or even potentially better performance for the same dollar than a lesser number of data center ssds?

Can I make the assumption that retail = consumer?

Here is my thought: the last 3 clusters I built with consumer SSDs did not survive burn-in. The reason I do the buyer's guide pieces is to get some of these lessons learned out there.

 

[RobertFontaine]

... Yes retail = consumer and I need to read that guide

 

[Patrick]

... Yes retail = consumer and I need to read that guide

Super! The consumer drives are a bit different. No PLP, and usually they are using the bleeding edge of low cost NAND. The enterprise drives have more OP, PLP and you still see MLC.

 

[spazoid]

Real data is much better than no data.

Used data center ssd's still seem to be more expensive than new retail ssds. I keep wondering whether lots of of little retail ssds in raid are a better value or even potentially better performance for the same dollar than a lesser number of data center ssds?

I bought 3x 250GB consumer SSD's that I've deployed in an AIO setup. They're configured as RAIDZ and accessed through NFS. In theory these disks should perform very well, but reality is quite different.
Don't get me wrong, my VM's run very well when not doing any intensive IO on the disks, but if I download something (1 Gbit internet service) while moving large files around, it will seriously slow down. I did not expect that, but I guess I underestimated DC class SSD's...

At the moment looking for DC class SSD's to replace this setup, but bargains are hard to come by in Europe...

 

[RobertFontaine]

I'm seeing decent pricing on the 3500's at this point. Is this the current sweet spot?

I've got a handf of desktop vms that I need to run concurrently for product testing and demonstrations and it seems like getting them each on their own physical ssd is the quickest path to
better performance.

Secondary question would I see equivalent improvement if I did a software raid 0 with the same number of ssd and put the vms on it...

Or Conversely should I spread the drives onto as many different sata channels/controllers as I can on the board?

Thanks as always
Robert

 

[MiniKnight]

I think one of my colleagues tried @RobertFontaine

His feedback was sequential read/ write is better with RAID 0 but IOPS not much better. The reason he did RAID 10 on the setup was because he wanted a bigger capacity pool and the sequential read speeds for the VM that needed it. Many times one of 50 VMs was doing heavy I/O so that VM got a lot more with RAID 10.

 

[RobertFontaine]

My use case so far is a bit more trivial. No matter how I configure it though I think I need one or two short stacks of not terribly big ssds to speed things up.
I have enough ram for the moment but I suspect I will be upgrading to 128gb next month as well. If that doesn't work it will be time to get off sata.