SandForce "SF-22XX" issues and workarounds

[Anvil]

There are a lot of variables.

I haven't decided on how to continue yet, I might start off by changing the multiplier back to 48x again and let it run until it's past the normal disconnect period (30-33 hours), or I might try keeping C states disables but reactivating the other settings.

I'm not sure about the PSU in the ASRock, it's 750W and it's not a cheap one, it's not Corsair though.
The other PSU's are mostly Corsair or Seasonic, 750-850W. (HX and/or AX)

[Christopher]

I have a noisy, mediocre, 500W OCZ Modular PSU. If I can put up with the awful noise, I could try it. Maybe SF2281s need bad power supplies, not good ones........ probably not, but if something changes, it'll be a place to start.

[sergiu]

Yes, I did it right
I verified that it was idling at the max non turbo multi, and performance power profile was set (with the change that drive sleeps: never)

I'll reboot and try again (for some reason, after I flashed back to the 2040 UEFI, the spash screen lasts for about 2/10s of a second.... I actually have to use a jumper to get into the UEFI now... seriously, power on to desktop is like 8 seconds now). Now the UEFI has been changed back to power savings (Cstates/eist/turbo/etc.) but the power plan is still high performance, but the CPU speed changes from 2.49mhz to 3.29mhz frequently (with a default 99.8 base clock, can't adjust it with anything other that whole increments, not tenths like many boards).

The speed change you described is the variation you see to the default speed in Mhz? or it is the real speed switching from 2.49Ghz to 3.29Ghz. If it is the last one, then EIST is still enabled. Another thing to concern... most PSUs from today are focusing on 12V line. I guess SSDs are usually taking power from 5V line, as it is closer to their needs. It might be the difference between SSDs as others might use 12V instead of 5V one.

[XII]

Don`t know if someone can reproduce it, but on my Vertex3-480 :
During setup, and Win install I had it on a Molex to Sata lead(no orange lead) and halfway the install, it shut down, several times.

Now hanging on a true Sata lead, it`s up and running for 76 hours.
(X58)

[Christopher]

The speed change you described is the variation you see to the default speed in Mhz? or it is the real speed switching from 2.49Ghz to 3.29Ghz. If it is the last one, then EIST is still enabled. Another thing to concern... most PSUs from today are focusing on 12V line. I guess SSDs are usually taking power from 5V line, as it is closer to their needs. It might be the difference between SSDs as others might use 12V instead of 5V one.

Yes, but the Seasonic X series uses DC to DC for the 3.3 and 5v rails.

Yes, EIST/Speedstep and Turbo are now enabled... but performance power mode is too. With all the power saving functions off, it was just sitting at the max non turbo ratio as it should.

I'd turn off TRIM if it didn't result in positively awful avg speeds. I'd be better off just living with the disconnects than running without TRIM. If I were having these problems but not trying to write 10+ TB a DAY, I'd not expect turning off trim to have much of an effect.

[sergiu]

Don`t know if someone can reproduce it, but on my Vertex3-480 :
During setup, and Win install I had it on a Molex to Sata lead(no orange lead) and halfway the install, it shut down, several times.

Now hanging on a true Sata lead, it`s up and running for 76 hours.
(X58)

That's a very important piece of information. Orange line is 3.3V which is probably most neglected by PSU manufacturers as from P4 days, most power is taken from 12V lines. I have not noticed until now that SATA power cable has 3.3v line and when I search for specs, I found that this line is optional and might not be implemented in all PSUs.

[Christopher]

Don`t know if someone can reproduce it, but on my Vertex3-480 :
During setup, and Win install I had it on a Molex to Sata lead(no orange lead) and halfway the install, it shut down, several times.

Now hanging on a true Sata lead, it`s up and running for 76 hours.
(X58)

During the Windows install, I think it's common since you can't turn off LPM (unless you have it in the BIOS).

You'll have to try it multiple times or just clone another drive to get Windows on the drive, then make necessary tweaks.

Maybe I should try it WITH a molex to sata connector. With my most successful setups with two motherboards, I can't get more than about 30hrs. If I try it with a new PSU or maybe just a molex to sata, something could possible change. I'm going to take a few hours to figure out how to proceed, the tonight I'm going to post my Mushkin Update of the Day, then swap some stuff around.

[sergiu]

Yes, but the Seasonic X series uses DC to DC for the 3.3 and 5v rails.

Yes, EIST/Speedstep and Turbo are now enabled... but performance power mode is too. With all the power saving functions off, it was just sitting at the max non turbo ratio as it should.

I personally trust a PSU only if it passes the oscilloscope test. I had a bad experience with a "good" PSU in the past. Also, to check if the issues are related to power saving features/ power lines, could you keep one or more cores from your CPU busy with a CPU torture program like Prime95 or any other one while running the endurance test? This should keep the power consumption relatively stable and allow for a much better regulation.

[Christopher]

I personally trust a PSU only if it passes the oscilloscope test. I had a bad experience with a "good" PSU in the past. Also, to check if the issues are related to power saving features/ power lines, could you keep one or more cores from your CPU busy with a CPU torture program like Prime95 or any other one while running the endurance test? This should keep the power consumption relatively stable and allow for a much better regulation.

I've been considering that very solution. Except that I was thinking at the 12 - 18 hour point I start Prime95. My system doesn't get louder (or much warmer) under prolonged CPU loads, so I won't be able to tell by listening (Which is how I like it -- quiet.) The Noctua U12 SE2 is highly recommended by me if you dig quiet. It's great for Sandy Bridge processors with just one fan at 800-900RPM, even with a substantial OC.

EDIT
Maybe I'll just try running P95 tonight as the Mushkin's only been running for a few hours now. If I can make it through the night, I'll just leave it until it A)disconnects or B) I get tired of it running P95 all the time (I was thinking all 4 cores, but one/two might be sufficient and leave me with some spare cycles for when I need to use the system.)

[XII]

During the Windows install, I think it's common since you can't turn off LPM (unless you have it in the BIOS).

You'll have to try it multiple times or just clone another drive to get Windows on the drive, then make necessary tweaks.

Maybe I should try it WITH a molex to sata connector. With my most successful setups with two motherboards, I can't get more than about 30hrs. If I try it with a new PSU or maybe just a molex to sata, something could possible change. I'm going to take a few hours to figure out how to proceed, the tonight I'm going to post my Mushkin Update of the Day, then swap some stuff around.

The Molex to Sata, is what gave me the trouble. Power comes from a AX1200, on a Sata now.
After the first reports on Anand, I sent him a mail, with a request if he could verify, but never got response.

[Christopher]

I had a bad molex to sata that I put in a friend's computer. Caused no end of problems, but the only reason I used it was because it was a really unconventional case style and made cable routing much easier. It would just cut out power to the drive, and this screwed up the Window's install. So I could see that happening. I don't see too many people using a molex to sata with a 480GB Vertex 3 though... probably not likely.

[XII]

I had a bad molex to sata that I put in a friend's computer. Caused no end of problems, but the only reason I used it was because it was a really unconventional case style and made cable routing much easier. It would just cut out power to the drive, and this screwed up the Window's install. So I could see that happening. I don't see too many people using a molex to sata with a 480GB Vertex 3 though... probably not likely.

It was a temporary setup, as I wanted to do it quick. To the Molex line, the casefans were connected, and a free one.
All Sata`s were used. Laziness bites your ass

[Christopher]

I'm thinking about just trying another motherboard; If it doesn't help the Mushkin's stability, at least I'll have another board to play with.

Thinking about the Maximus IV Gene-Z... a lot.

I guess in about 20hrs I'll know whether I'm currently making any progress as it stands now, so I suppose I can wait to place an order until tomorrow night/Monday morning.

[Anvil]

I ended up letting it idle for 12 hours, no change at all to WR Delta, power on hours still looks to double while idling?

It is now running on the X58 with the 11.x orom and the alpha RST 11.xx as well.
Power savings are disabled on this rig so if it works out it should go well past 33 hours w/o issues.

WRD has gone from 57 to 56 in 4 hours so it's still decreasing.

[Christopher]

It looks like someone was able to get a Maximus IV Gene-z BIOS from an Engineer at ASUS which had the 11.x in it, and I'm trying to figure out if there are some tools you can use to add it to UEFIs, or if it has to be a regular BIOS. I'm trying to sift through about 900 forum posts to figure this out, so it might take a while, but I think I'm going to order the board anyway. Of course, if the 11.x orom could just be inserted into any UEFI, I could just use the boards I've already got laying around. If anyone knows a way to put 11.x into an Intel DP67BG UEFI, that would be supremely cool.

EDIT
After the last changes I made, I'm already sitting at 30.5 hours without crashing. In another 90 minutes it'll be a record for me, so that's positive.

[scorp]

Haven't posted here in quite a while

I have an OCZ V3 ( with stability problems ofc ) and I've noticed that E9 has skyrocketed since I've installed the latest FW ( 2.13 ). In fact it's now bigger than F1, while ~10 days ago ( when I was on 2.11 ) it was at 87% of F1. Has anyone else noticed a similar behavior in 1.3.2 from Corsair or in 2.13 from OCZ ? ( I'm not sure if any other manufacturers released newer FWs lately ).

[Christopher]

Haven't posted here in quite a while

I have an OCZ V3 ( with stability problems ofc ) and I've noticed that E9 has skyrocketed since I've installed the latest FW ( 2.13 ). In fact it's now bigger than F1, while ~10 days ago ( when I was on 2.11 ) it was at 87% of F1. Has anyone else noticed a similar behavior in 1.3.2 from Corsair or in 2.13 from OCZ ? ( I'm not sure if any other manufacturers released newer FWs lately ).

So your V3's writes to NAND (E9) have gone up substantially, so much that actual writes to NAND are now outstripping host writes(F1)???

I don't think that's possible. In a worst case scenario where you workload is 100 percent random, F1 and E9 should be the same and increase at the same rate. You can't have more writes to NAND than you have in host writes. Most of the time writes to NAND should be between ~75 to ~85 percent of host writes. I could see why you'd be alarmed though...

Unless your workload has some crazy amount of write amplification, and that amplification shows up in NAND writes (I don't think it does), I don't see how it's possible without some sort of error. With SF, your WA should be well below 1 on a client system, and in reality it should be around .7 with Windows.

Can you post a screen shot from CrystalDiskInfo?

[Anvil]

There are no such issues with my 1.3.2 FW.

I have seen the issue though, on an early firmware and in my case it was on a Vertex 2.

I'd give it a few days and reflash the drive if it doesn't change.

I'm not sure that 2.13 = 1.3.2 but it's very likely.

[scorp]

My work pattern hasn't changed at all. The PC is very lightly used ( mainly browsing with Chrome, reading emails with Thunderbird; all the games are installed on another SSD / HDD ). It's not that I'm really worried, but I think this might be some kind of a "fix" for the stability issues, a bug in the FW ( probably affecting the reporting ).

[Ao1]

I don't think that's possible. In a worst case scenario where you workload is 100 percent random, F1 and E9 should be the same and increase at the same rate. You can't have more writes to NAND than you have in host writes.

Try running uncompressible 4K random writes and you will soon see that NAND writes can exceed host writes WA & WL can offset savings in compression. For most people the value will be below 1, but with certain workloads it can be higher. (I've seen as high as 3.3)

[scorp]

It's entirely possible that I get small writes on the drive. This morning I had E9 at 232 and F1 at 227. In the meanwhile E9 got to 235 and F1 to 228, hence writting ~3G on the NAND for only 1G of actual data.

[Christopher]

Lets hope it works out and is reproducible!

The reproducible part is what I'm aiming for...

But at this point, I have a SF2281 which will disconnect like clockwork every 30 hours (or I can get it to disconnect earlier). If.. and I say again... IF I can get the Mushkin to:

A) Remain stable for a few more days with the changes I've made (I'm well past my previous record)

AND

B) Get a disconnect with the new motherboard I have on the way on the same time frame (as the P67 and H67 boards)

THEN

C) I think I'll have something interesting going on here.


All power saving and turbo technologies are enabled, current power draw at the wall is 47w.

Using RST 10.6 drivers, Intel DP67BG with 2040 UEFI revision.

Unfortunately, even if (and that is a big IF) I can get the drive to be stable, I'm worried that I might have also changed something without knowing, and won't be able to reproduce the error. So if the drive crashes soon I don't have that problem. If it doesn't, then how long should I wait? 60 hours? 80?

[Christopher]

It's entirely possible that I get small writes on the drive. This morning I had E9 at 232 and F1 at 227. In the meanwhile E9 got to 235 and F1 to 228, hence writting ~3G on the NAND for only 1G of actual data.

Are you sure your drive is properly aligned? If you run AS SSD, does it say "1024K OK" in green text (or some multiple of 1024 - OK)?

If the drive is sector aligned and not 4K aligned, this could account for the write amplification if that's what is going on.

AS SSD download page (it is at the bottom)
http://www.alex-is.de/PHP/fusion/dow...&download_id=9

[squick3n]

Are you guys deploying either of the Toshiba Toggle NAND drives (Wildfire and MAXIOPS)? I had all sorts of issues with the vanilla Vertex 3, but the MI drive has given me no issues. Obviously that's nothing scientific, and since OCZ is the only vendor using LTT (right?), there's no way to test it.

[Christopher]

Are you guys deploying either of the Toshiba Toggle NAND drives (Wildfire and MAXIOPS)? I had all sorts of issues with the vanilla Vertex 3, but the MI drive has given me no issues. Obviously that's nothing scientific, and since OCZ is the only vendor using LTT (right?), there's no way to test it.

I'm using the Mushkin Chronos Deluxe 60GB, which is like a 60GB Wildfire or MaxIOPS. 60GB of Toshiba 32nm Toggle NAND. Anvil is using the Corsair Force 3 120GB with asynchronous Micron 25nm.

With respect to lifetime throttling, it isn't clear who is using LTT and who isn't, except to say that Mushkin and Corsair don't use it, so I think that all the drives using reference SF FW do not have LTT. OCZ may or may not still use LTT, but I have no way of knowing.

Mushkin, Patriot, OWC, Edge, and a few others all have their drives assembled in America and all use the same firmware. A Patriot Pyro and a Mushkin Chronos vary only by the label of the front of the case.

I have heard that the Toggle equipped drives were better in this respect, but people have been saying a lot of things about SandForces. OWC switched their 2281 Synchronous equipped SF SSDs to Toggle NAND and are using the 2282 processor in the 240GB version.