SSDs stability under 24/7 100% i/o load

[Lapsio]

I'm using crappy Intel NUC with Celeron as home hypervisor for VMs. Obviously it's under 100% load 100% of the time because it's complete garbage. It has 8GB of ram (max supported) and 32GB of swap which is almost constantly used above 80% since 8GB of RAM is almost nothing for VMs hosting. Due to heavy load caused by VMs SSD in NUC is almost constantly under 100% load both because of actual i/o and because of intense memory swapping. I'm running 8 virtual servers on this host.

 

I used to use 120GB GoodRAM CX200 SSD as drive and it was... garbage, like whole NUC. But it worked. Quite stable actually. Unfortunately due to growing amount of snapshots I ran out of space and had to replace SSD. In the past I had really bad experience. I had pretty bad experience with Samsung SSDs under prolonged 100% load (even PRO) that were behaving quite unstable (comparing to that CX200 at least) and I decided that I'm not gonna spend 200$ on SSD for computer that had cost me like 150$ total for everything else so I went with GoodRAM IRDM 240GB.

 

Unfortunately IRDM 240GB crashes under heavy load in similar manner to Samsung drives. After around half an hour of 100% i/o load during for example filesystem balancing whole SSD craps its pants and crashes with ton of I/O errors and basically whole system crashes and motherboard doesn't detect SSD until NUC is completely turned off. When I added excessive cooling by placing NUC on top of 140mm fan running at 100% rpm it worked more stable but come on, it's not really acceptable solution.

 

So my question is - what SSDs are more suitable for prolonged 100% 24/7 load? Preferably not server SSDs since those can cost more than thousand of NUCs...

[RAM555789]

I would recommended going with a performance HDD like WD Black lineup but their are enterprise solutions for SSDs that might work for you.

https://www.newegg.com/Product/Product.aspx?Item=1Z4-002P-00GF6

[KarathKasun]

No SSD is suitable for 24/7 writes.

 

Get a better system for a server or reduce load on what you have.

[unijab]

What's your budget?

[unijab]

You should solve the swapping issue (either by limiting ram usage by your VMs ) or by adding more RAM.

 

Then re-evaluate where you are with your disk i/o

[Lapsio]

1 hour ago, unijab said:

You should solve the swapping issue (either by limiting ram usage by your VMs ) or by adding more RAM.

 

Then re-evaluate where you are with your disk i/o

This NUC doesn't support more on motherboard/CPU level. It's max supported unforutnately.

 

2 hours ago, unijab said:

What's your budget?

Potato and 3 bottle caps. On more serios side - like I said I paid around 150$ for NUC and around 300$ for whole machine including GoodRAM CX200 120GB, Kingston HyperX 8GB RAM and external 2TB Seagate IronWolf NAS HDD in USB 3.0 enclosure for shared NAS storage (attached to VMs via NFS over localhost) and junk like USB hubs etc. I'd like to stay within "rational" budget. By "rational" I mean that if I wanted to spend 300$ on SSD alone I could just as well replace NUC with something stronger instead. So I guess something like... dunno 100$? Maybe 150$ top. I'm rather looking for around 240GB since that 120GB was only slightly too tight.

 

21 hours ago, KarathKasun said:

No SSD is suitable for 24/7 writes.

 

Get a better system for a server or reduce load on what you have.

SAS SLC enterprise SSDs are usually write-optimized. Especially ones made for SAN caching etc. But yeah I'm afraid consumer SSDs may not be best choice for write intensive tasks in terms of durability. That said multiple tests have shown that it's not that bad actually and consumer SSDs can easily handle hundreds of TBW without failure.

 

I returned that IRDM and replaced with IRDM PRO 480GB for now (240GB variant seems to be mostly out of stock). It has metal housing so maybe It won't overheat that easily. I'll check soon if it handles load better than that poor plastic toy...

[Lapsio]

22 hours ago, RAM555789 said:

I would recommended going with a performance HDD like WD Black lineup but their are enterprise solutions for SSDs that might work for you.

https://www.newegg.com/Product/Product.aspx?Item=1Z4-002P-00GF6

TLC is really bad idea for write intensive tasks. Especially if SSD will be filled up in high degree (which strongly degrades TLC performace). That's why I went with GoodRAM IRDM in the first place - they're MLC. Just like Samsung PRO series (just cost less kidneys per GB).

 

Unfortunately that lower cost comes at some compromises. I just got info from Wilk Elektronik support that those drives don't have for example temperature sensors (I asked about that specifically since S.M.A.R.T of IRDM reported constant 33 degree temperature which obviously was bullshit as probably ambient temperature in this NUC is already above 33 deg). Which is really unfortunate in my case since cooling of that NUC is terrible joke.

[KarathKasun]

Even at 1pb of writes, remember that you can possibly write 500mb/s+ to a SSD.  At 100% writes, this is 55h of use.  With a 50/50 read/write split it gives you 110h of use.  If you drop the load to half of that, its 100 days of use.

 

A petabyte sounds large, but at the write rates of SSD's, that amount of writes goes by in the blink of an eye.

[Neftex]

maybe the temps are the problem? can you leave the ssd out and blast it with the fan?

[KarathKasun]

2 hours ago, Lapsio said:

This NUC doesn't support more on motherboard/CPU level. It's max supported unforutnately.

This means that it is not suited for the task you are using it for.  Get different hardware.

[Lapsio]

2 hours ago, Neftex said:

maybe the temps are the problem? can you leave the ssd out and blast it with the fan?

it was indeed issue. It was related partially to fact that normal IRDM was plastic and NUC has weird metal frame around disk that seems to partially work as heatspreaded that obviously doesn't work when SSD is plastic.

 

I repeaded tests and IRDM pro doesn't seem to fail and temperatures stay below 65 deg under load.

 

3 hours ago, KarathKasun said:

Even at 1pb of writes, remember that you can possibly write 500mb/s+ to a SSD.  At 100% writes, this is 55h of use.  With a 50/50 read/write split it gives you 110h of use.  If you drop the load to half of that, its 100 days of use.

 

A petabyte sounds large, but at the write rates of SSD's, that amount of writes goes by in the blink of an eye.

No. at 550 MB/s rate you get 45 TB/day (550*60*60*24/1024/1024) which gives you 16PB / year. But yes, it's mixed load so significantly less plus 550 MB/s is for sequential load. Operations like swapping and VMs I/O are highly random so their performance is significantly worse. It took me around 40 minutes to balance 90GB of data on this IRDM so it gives around 40 MB/s of heavily random mixed I/O. CPU wasn't maxed.

[unijab]

3 hours ago, KarathKasun said:

Even at 1pb of writes, remember that you can possibly write 500mb/s+ to a SSD.  At 100% writes, this is 55h of use. 

 

Math is a little off according to google

 

[unijab]

4 minutes ago, Lapsio said:

swapping and VMs I/O are highly random so their performance is significantly worse.

 

4 minutes ago, Lapsio said:

so it gives around 40 MB/s of heavily random mixed I/O.

Sounds like you need a 800p or 900p intel ssd...

but those are super expensive.

[Lapsio]

4 minutes ago, unijab said:

 

Math is a little off according to google

 

it sounds okay. 1 pb for 23 days is around 16 pb per year since 365 / 23 ~ 15.8 which is around 16pb / year like I said.

[Lapsio]

4 minutes ago, unijab said:

 

Sounds like you need a 800p or 900p intel ssd...

but those are super expensive.

That NUC doesn't have M.2, only single SATA

 

But yes I considered 900p for workstation since it sounds like perfect drive for large swap for more serious VMs hypervisor.

[KarathKasun]

1 hour ago, unijab said:

 

Math is a little off according to google

 

Missed a zero, ~555hrs vs ~55.5.

 

Even then MLC flash 240gb drives are only good for ~1pb of writes last I had checked.  So at 40mb/s you get under a year of usage, 289 days.  120gb drives would have around half the write endurance.