3 ms·
Random bit flips is a much longer time frame than we're currently seeing. Also, any services should be restartable at any time, so the only random bit flips tha
by naasking 7y ago
Random bit flips is a much longer time frame than we're currently seeing. Also, any services should be restartable at any time, so the only random bit flips that should matter are those in the kernel itself. The kernel data should be significantly smaller and so less likely to have a bit flipped.
If uptime and availability become a serious concern, then a partitioned system in which kernel data is in ECC memory while the remainder runs on cheaper non-ECC memory is possible as well.
- dijit 7y agoI think you greatly underestimate a few things; one being the windows kernel size in memory, the other being the sheer volume of bit flips in high density memory. A single 32GB SODIMM (as I have in my laptop, because it's a workstation laptop) see's roughly 12 bit-flips a month, these are correctable. This is pretty much in line with the 6-or-so bit flips I see on ECC memory inside servers (of which I have a few thousand, each having 32G modules spread over more area and enclosed in a heavy chassis, thus, less flips). Bit flips are so common that it's even possible to have "double bit flips" in the same page, which is uncorrectable even with ECC. So I think you grossly underestimate the percentage of bit flips, most are silent though (or attributed to bugs in software) so I understand you don't see it. If you had ECC memory you would. My other point about Windows kernel being fairly large in terms of surface area is relatively self-evident, but I'll come back with some solid numbers, because those thousands of servers I was mentioning are all Windows. EDIT: As promised I checked the non-application memory usage of my fleet and the rounded value of Windows memory footprint by itself is 1.89GiB. FS Cache is also excluded.
- naasking 7y ago> I think you greatly underestimate a few things; one being the windows kernel size in memory [...] As promised I checked the non-application memory usage of my fleet and the rounded value of Windows memory footprint by itself is 1.89GiB That number is frankly absurd. No doubt it includes all sorts of things that shouldn't be included in a proper microkernel, which just goes to show to all those people that call the NT kernel a microkernel (or close to it), are full of it. In principle, some book keeping memory for processes and the page tables are roughly all you need to live in kernel space.
- tracker1 7y agoIs that actual memory usage, or driver/device addressing in kernel space?
- ronsor 7y agoThere's no way the windows kernel is 1.89GiB. That's ridiculous.
- dijit 7y agoSome things to note when talking about windows though; nearly their entire graphics stack lives in the kernel. They also have multiple environment subsystems running at the same time (POSIX, OS/2 and Win32) which, are running constantly and unavoidable, so I count them as kernel. In addition they have things like the “security reference monitor” which resides entirely in the kernel. It’s really apples to oranges, NT is truly heavy especially on server 2016 (which is where I grabbed the numbers) I have comparatively fewer 2012r2 machines but I can aggregate the numbers for those too if you like.
- luckman212 7y agoCurious (and genuinely concerned) how does one get a reading or otherwise measure these bit flips?
- dijit 7y agoThis article will tell you how to do that. Assuming that your edac works on your chipset. http://www.admin-magazine.com/Articles/Monitoring-Memory-Errors http://www.admin-magazine.com/Articles/Monitoring-Memory-Err...