Intel510 slows way down.

[happy_gopher]

I am running an Intel 510 with IOMeter 1.1.0, 100% sequential Writes. I start the test, and this drive will run balls to the wall for about 16 hours, then slows down to about 30MB/s over the course of an hour or less. It has done this to me every time I have started writing to it.
Its new, maybe has 60TB written to it at most. It uses the Marvel controller that I know nothing about.
I will flash the firmware as soon as I find a new version.
(I have an X25-E and it is worn out (516TB Written), and does something similar, but this drive is new...)


Any ideas?
Every seen this before?
I was thinking my access spec, but I have messed with it alot! No Luck!D:

Hg

 

[HendrixFan]

I am running an Intel 510 with IOMeter 1.1.0, 100% sequential Writes. I start the test, and this drive will run balls to the wall for about 16 hours, then slows down to about 30MB/s over the course of an hour or less. It has done this to me every time I have started writing to it.

Are you really running a write test for 16 straight hours?

 

[VirtualLarry]

Are you really running a write test for 16 straight hours?

Yeah, WTF! Sounds like the drive ran out of free blocks.

Edit: On second thought, I really don't know what's going on. The drive should have ran out of free blocks before 16 hours.

 

[Zap]

Hello happy_gopher, and welcome to AnandTech Forums.

Any ideas?

Yes. You are hammering your SSDs until you wear them out. You are intentionally killing your SSDs. Seriously, don't you know that an SSD has a limited lifespan for writes?

Once every bit of the drive is written to, the drive has to clean up deleted blocks before it can be written to again, AKA Garbage Collection. This is what Trim does (by prompting the drive to clean up during idle time). The Marvell controller tries to keep performance up by putting off Garbage Collection for as long as it can. Normally this works very well. By "normally" I mean for people who use their computers for their daily tasks such as web browsing, playing back media, playing games, doing a bit of work, etc.

If you read the sticky you will find this gem:

THINGS TO NOT DO ON AN SSD

Do NOT defrag an SSD.
...
Do NOT endlessly run benchmarks on an SSD.
...
Do NOT endlessly run anything that writes huge amounts of useless data to your SSD.

EDIT: I see from your other posts that you are testing SSDs for work. What is the goal of all that testing? Like I said before, the Marvell controller works best with Trim.

 

[happy_gopher]

Hello happy_gopher, and welcome to AnandTech Forums.



Yes. You are hammering your SSDs until you wear them out. You are intentionally killing your SSDs. Seriously, don't you know that an SSD has a limited lifespan for writes?

Once every bit of the drive is written to, the drive has to clean up deleted blocks before it can be written to again, AKA Garbage Collection. This is what Trim does (by prompting the drive to clean up during idle time). The Marvell controller tries to keep performance up by putting off Garbage Collection for as long as it can. Normally this works very well. By "normally" I mean for people who use their computers for their daily tasks such as web browsing, playing back media, playing games, doing a bit of work, etc.

If you read the sticky you will find this gem:


EDIT: I see from your other posts that you are testing SSDs for work. What is the goal of all that testing? Like I said before, the Marvell controller works best with Trim.

Yep, my goal is to break them. Some mfg's will give you a TBW that the drive is capable of, some just stick with the 5K writes for MLC, 30K for E-MLC, 100K for SLC. So my goal is to validate those lifecycles by writing at maximum speed until they fail.
What I am finding, is that MLC is underrated in many cases. I have a 40GB MLC that should crap out at 200TB writes. (40GB X 5K writes). Today, it blew past 645TB writes. Pretty amazing right? That means I am approaching 50% E-MLC territory, and 231 (SSD Life Left) is only down to 66. Now, I have not come anywhere near the life of E-MLC/SLC drives, but I have alot of time to get there. My good E-MLC 200GB useable drive should reach 6PB this year, which is the supposed life of the drive. Cant wait to see that! My SLC drives are supposed to be good for 20+PB, a shocking amount of data!
I havent figured out how to manipulate TRIM for my intended purposes yet, but if TRIM is the problem, ie i am filling the drive, and trim is throttling it to deal with garbage collection, then it will need some tweaking.

 

[Tsavo]

I am running an Intel 510 with IOMeter 1.1.0, 100% sequential Writes. I start the test, and this drive will run balls to the wall for about 16 hours, then slows down to about 30MB/s over the course of an hour or less. It has done this to me every time I have started writing to it.
Its new, maybe has 60TB written to it at most. It uses the Marvel controller that I know nothing about.
I will flash the firmware as soon as I find a new version.
(I have an X25-E and it is worn out (516TB Written), and does something similar, but this drive is new...)


Any ideas?
Every seen this before?
I was thinking my access spec, but I have messed with it alot! No Luck!D:

Hg

If that Marvell controller is anything like a SF controller, it'll neuter driver performance in an attempt to retain drive lifespan.

 

[Zap]

What I am finding, is that MLC is underrated in many cases.

This is true. Manufacturers don't want any returns within the warranty period.

 

[Old Hippie]

Yep, my goal is to break them

Cool!

Could be very interesting.

 

[=Wendy=]

Warranty will not cover burned out NAND, unless the smart data returns a value that does not meet the specifcations of the said NAND.

Although I would be very interested to hear how many P/E cycles MLC NAND can withstand.

 

[Bauss]

Well the way I see it, you've essentially taken the SSD to the point where there's literally no where else to write, and you've written faster than the IGC's ability to keep up. Assuming the 510 can write ~200MB/s to the drive over that time, you've written ~12TB to the thing over 16hrs-- WAY more than should be written to it in the regular usage scenario that these things were designed for.

To be honest, I doubt you'll ever find an SSD with garbage collection that can keep up with what you're trying to do. It's like taking a Civic to the 24hrs of Le Mans, and wondering why it couldn't finish...

 

[n7]

I would love if you'd share more results of this as you get them...would be interested in hearing which drives are lasting longer than they are rated, which less, etc.

 

[groberts101]

yeah.. many are right on target here. Firmware has no where near the capability yet to keep up with the necessary housekeeping for that constant dataset. Not to mention that Trim/GC comes at the cost of increased WA so they will never allow it on consumer products even if it could be adjusted to keep up. SLC isn't hit with that erase before write penalty whereas there's just no way around it with MLC.

If the mfgrs could adjust firmware to compensate?.. they surely would as this speed degradation problem is across the board and regardless of the controllers used.

All I can say is that if you think that's bad?.. try to do that with a Sandforce based drive. The controllers are purposely designed to slow down when hit like that with Durawrite throttling. I would be extremely interested to see how much data could be written to one with the unique life-span saving compression algorithms built into it though.

Kinda cool to see this testing anyways as it's surely a 3 pointer for all the "just use it and don't worry about lifespan" advocates out there. :thumbsup:

 

[happy_gopher]

I would love if you'd share more results of this as you get them...would be interested in hearing which drives are lasting longer than they are rated, which less, etc.

New data.
I just had my second 40GB Vertex2 failure.
SMART 241 says 759TB. (the last one made it 741TB)
Failure mode: half speed writes. 101MB/s max
This is a wierd issue, still no cells retired, no other problems.

But half speed is bad for me, so I consider it dead.

 

[nanaki333]

wow.. so it's safe to say once you hit the 700TB mark, your drive is on it's way out and should be decommissioned. it's a ticking time bomb by then.

what are they rated at from the factory?

 

[Zap]

what are they rated at from the factory?

From his earlier post:

I have a 40GB MLC that should crap out at 200TB writes. (40GB X 5K writes).

 

[nanaki333]

From his earlier post:

damn... if i only i went back and read it as a refresher.... it's been awhile since his last post ok!?

 

[PeeluckyDuckee]

Ssd's are still expensive for the raw performance one gets, with deteriorating performance and shortened life span. I rather buy bulk mechanical drive on the cheap and not really have to worry about its write life span. Its cheap enough i dont even have to think twice about dumping it for a bigger denser drive down the road. The speed of a ssd is nice, but its a very expensive consumable.

 

[jiffylube1024]

I am running an Intel 510 with IOMeter 1.1.0, 100% sequential Writes. I start the test, and this drive will run balls to the wall for about 16 hours, then slows down to about 30MB/s over the course of an hour or less. It has done this to me every time I have started writing to it.
Its new, maybe has 60TB written to it at most. It uses the Marvel controller that I know nothing about.
I will flash the firmware as soon as I find a new version.
(I have an X25-E and it is worn out (516TB Written), and does something similar, but this drive is new...)


Any ideas?
Every seen this before?
I was thinking my access spec, but I have messed with it alot! No Luck!D:

Hg

You should state your purpose in your post - are you deliberately trying to test the manufacturers' rated life claims?

I'm absolutely amazed that it goes 16 hours before slowing down! I cannot believe that is even possible, is that all because of TRIM that it's able to last that long?

You're providing a valuable service by posting these results using real, store-bought drives, and not some possibly cherry-picked drives given to reviewers. I doubt you will find too many people on these boards who have run into your issue, though, as the vast majority of us do not run write torture tests to our drives, because we want them to last for years!
---

I like how you post about your drive "it's new - only has [up to] 60TB written to it" -- assuming you've got a 128GB drive, 60TB is already almost 500 P/E cycles on there, and these are rated for what, 3000-5000 cycles?