Microsoft hardware analysis of crashes on a million PCs

[SafeHarbor]

Here's something interesting from Microsoft research. They statistically analyzed automated post-restart crash reports and mini-dumps for a million consumer PC devices for an eight-month period. They found that laptops, in general are more reliable than desktops, having more RAM leads to more crashes (!), and underclocking substantially improves reliability while overclocking unsurprisingly does not. Also big manufacturer boxes are more reliable than white boxes.

Report is here: https://research.microsoft.com/pubs/144 ... ingale.pdf

What they didn't mention that I'm still a little curious about is whether turning on speed-step or whatever it's called on my desktop and my Thinkpad is beneficial. I've watched CPU speed go up and down, up and down continuously on my C2D laptop and AMD quad-core desktop, and I've always suspected that that the speed-regulation process just can't be very good for it. I keep both set for "Maximum performance" anyway (when plugged in) because I'd rather have all the gitty-yup instantly available.

[RealBlackStuff]

They didn't mention Speedstep as such, but they did mention that lower-than-rated CPU speeds (whether induced by Speedstep et al, or via underclocking) improves reliability considerably.
By switching off Speedstep (in your case: run at full speed all the time), I think you do your CPU (and yourself) a disservice.
Maybe you should keep your "full-speed" settings, but add underclocking into the mix, to get the best of both worlds.

[dr_st]

Thank you for the interesting article.

What they didn't mention that I'm still a little curious about is whether turning on speed-step or whatever it's called on my desktop and my Thinkpad is beneficial. I've watched CPU speed go up and down, up and down continuously on my C2D laptop and AMD quad-core desktop, and I've always suspected that that the speed-regulation process just can't be very good for it. I keep both set for "Maximum performance" anyway (when plugged in) because I'd rather have all the gitty-yup instantly available.

I wonder what makes you think that it can't be good for it. Have you seen any evidence of problems caused by it? You say yourself that they reported nothing of the sort in this research.

[SafeHarbor]

Well, here's one late answer... Actually I have the Nvidia graphics in the T61p and there's a theory that build defects and heat/use will eventually kill it. Rather than having SpeedStep continuously yanking the temps up and down, a smooth continuous temperature range along with conscientious use of TPFan may just lengthen the time before failure. That said, this is a 2007-made TP, so it's right in the doomed range. Then again, it may continue working for years. Other than that, the 61p has been a favorite of mine since I got it.

Lamar

[Navck]

Funny little thing about RAM induced kernel panics: Why do you think servers with massive amounts of RAM (Read: 96GB and up) only use ECC RAM?

Think about it this way: For every bit of memory you have, you have a 1 in some large chance of it getting corrupted. If you have a lot of bits... That number starts closing towards "significant" for your memory operations. ECC lets you have more tolerance to memory corruption but regular RAM combined with your memory controller does not like to have too many unrecoverable errors. Say: More than a single bit.

Annnnd throw in some cosmic rays, nearby interference (And possibly on board interference) plus all sorts of fun. Watch as your system throws up on you. Nothing new here, except I'm sure some people will scream at me to replace my RAM and CPU with NAND and everything will become magical superfast and better, since known concepts like this (Why else do servers use ECC RAM given the amount they have?) have been around forever. (Unlike shrinking feature sizes on NAND which decrease your rewrite cycle life. Funny thing: Decreasing feature sizes for your memory modules will also raise your unrecoverable error probability per bit, at the same time it increases density... Fun things.)