new xeon v4 cpus

[peterguthrie]

uninstalled all windows 10 'metro' apps, still got the runtimebroker issue. ONLY happens on the machines in the office with more than 64 cores. I reckon it is a windows 10 issue that needs to be fixed. Hopefully once more people get these machines microsoft will fix it.

[bazuka]

Hi everyone,

im also expecting cpu's (2x40 = 80cores) to arrive during this week, and i just saw a problem here with vray and not detecting the total amount of cpus,

friend of mine already got them and he has the same problem with vray, thing is he tested corona, and everything is working great, so Peter my advice to you to try corona and see is everything working as it should...

corona test



vray test



according to this i think the problem is with vray, sorry vlado

waiting for updates...

p.s.

before these 2x40 cores, friend had 2x28cores and vray was using all cpu's so i guess there is a limit to lets say 64cores or something like that, perhaps?

cheers

[benb32]

Ok Bazuka, a shame you haven't read this thread though - the experiences here are far more to do with Windows 10, rather than vray, not behaving with large core count xeon... Also vray has supported greater than 64 cores for a while now (previously with an environment variable, and since 3.3 by default).

[vlado]

im also expecting cpu's (2x40 = 80cores) to arrive during this week, and i just saw a problem here with vray and not detecting the total amount of cpus,

Sorry, Bazuka, there's no problem with V-Ray. For the 3.30 builds, you need to enable support for more than 64 cores for V-Ray by setting the environment variable VRAY_USE_THREAD_AFFINITY to 1. This is enabled by default for later builds.

Best regards,
Vlado

[bazuka]

Hi vlado, he did enable the env var, and still not working,

but we found out whats the probl, according to windows, i think u should better read this, you will understand, but ill try to explain to other people with simple words whats going on

https://msdn.microsoft.com/en-us/lib...or=-2147217396
https://blogs.msdn.microsoft.com/b8/...al-processors/

so someone from Microsoft made that if u have (computer with) more than 64cores, windows will make a 2 groups! (dont know who was that smart to make this so stupid, no case in linux)

group1 (with max 64cores) and group2 (other max 64cores group), so people like a friend of mine with 80cores has also 2 groups of cpu's, divided by 40cores in each group (since they are physical cores) and as i explained vray see only that one group of the 2 groups...(that is 40cores)

soooo, the only solution is to, i dont know, rewrite vray or something so it could read both of the groups, thing is software should be written to read groups, (looks like guys from corona did that, or what ever), just to make it clear here, im not saying that corona is better, thing is we tried with houdini 15 and had the same problem, h15 see only one group ->40cores...

article about houdini problem
http://forums.odforce.net/topic/2353...ount-machines/

but as far as i see even in Peter case still not working, right? with that env variable?

[bazuka]

vlado how exactly vray see cpu amount? how do u detect number of cpu's/cores

as u can see there r groups of cpus

[vlado]

vlado how exactly vray see cpu amount? how do u detect number of cpu's/cores

We are using the Windows API for working with processor groups that was introduced in Windows 7:

			int useThreadAffinity=getUseThreadAffinity();
			if (useThreadAffinity) {
				// try using groups-aware API available in Windows 7 onwards
				TGetLogicalProcessorInformationEx pGetLogicalProcessorInformationEx =
					(TGetLogicalProcessorInformationEx) GetProcAddress(GetModuleHandle(TEXT("kernel32.dll")), "GetLogicalProcessorInformationEx");

				if (NULL != pGetLogicalProcessorInformationEx) {
					const LOGICAL_PROCESSOR_RELATIONSHIP relation = RelationNumaNode;
					DWORD bufferLen = 0;
					pGetLogicalProcessorInformationEx(relation, NULL, &bufferLen);

					VUtils::UPtr< SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, VUtils::DeleterWithFree< SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX > >
						buffer((PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) malloc(bufferLen));

					if (pGetLogicalProcessorInformationEx(relation, buffer.get(), &bufferLen)) {
						const DWORD recordSize = buffer->Size;
						const DWORD nodeCount = bufferLen / recordSize;

						const BYTE *it = reinterpret_cast< const BYTE* >(buffer.get());
						const BYTE *itend = it + bufferLen;
						int prefixSum = 0;

						for (DWORD i = 0; it != itend; it += recordSize, ++i) {
							const SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX &ri = *reinterpret_cast< const SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX* >(it);
							NodeInfo nodeInfo;
							nodeInfo.group = ri.NumaNode.GroupMask.Group;
							nodeInfo.mask  = ri.NumaNode.GroupMask.Mask;
							ULONG dummy = 0;

							computeNumaNodeResources(nodeInfo, prefixSum, dummy);
						}

						numProcessors = prefixSum;
						return;
					}
				}
			}
//...

/// compute cpu mask and prefix sum of available logical processors for a NUMA node
/// @param [in, out] nodeInfo record containing node info; mask and prefix-sum fields to be updated by this routine
/// @param [in, out] prefixSum running count of accessible logical processors
/// @param [in, out] processNodeCount total number of NUMA nodes available to this process
/// @param processMask custom processor mask for this process
template < typename T >
static void computeNumaNodeResources(
	NodeInfo& nodeInfo,
	T& prefixSum,
	ULONG& processNodeCount,
	const Mask processMask = Mask(-1)) {

	// if the intersection of process and node masks is non-empty, we can run on this node
	const Mask nodeMask = nodeInfo.mask & processMask;

	if (0 != nodeMask) {
		prefixSum += T(VUtils::__popcnt((size_t) nodeMask)); // count available processors so far
		++processNodeCount;
	}

	nodeInfo.mask = nodeMask;
	nodeInfo.prefixSum = prefixSum;
}

Best regards,
Vlado

[bazuka]

just an update , looks like its working....

friend just downloaded a latest ver 3.3.05 and env variable , my bad, im so sorry vlado , friend was using some older build of vray and he didnt told me that, and we were trying to use the env variable with previous build



once more, sorry

best regards

[vlado]

No worries, glad that it's working now

Best regards,
Vlado

[EL_CISNE_POST]

I'm actually having the same problem with the affinity. I'm using a new dual xeon v4 and most of the times (with my older machine) I use to keep working in simple stuffs while the backburner server is on, so I use to set the affinity to something that let me work without problem, but with this new machine, the computer almost collapse everytime the render starts, then I'm able (with a lot difficulty) to set the affinity, but when a frame is over, it takes again all the cores again, so, the machine collapse until I set the affinity again... is insane, and I was wondering if I was the only one with this issue. Anyone found a solution for this?

[peterguthrie]

a short update for anyone else using windows 10 and more than 64 cores... I'm very happy to report that the windows 10 anniversary update fixes the start menu and runtime broker issues. Still can't change affinity but learning to deal with that.

[vlado]

I'm actually having the same problem with the affinity. I'm using a new dual xeon v4 and most of the times (with my older machine) I use to keep working in simple stuffs while the backburner server is on, so I use to set the affinity to something that let me work without problem, but with this new machine, the computer almost collapse everytime the render starts, then I'm able (with a lot difficulty) to set the affinity, but when a frame is over, it takes again all the cores again, so, the machine collapse until I set the affinity again... is insane, and I was wondering if I was the only one with this issue. Anyone found a solution for this?

Affinity was never meant to solve this problem; the correct thing to do is set the priority of the 3ds Max process to "Below normal" or "Low". You should never tinker with the CPU affinity of a process.

Best regards,
Vlado

[Sbrusse]

The trick is to set the server.exe from backburner to low, that will force the resulting max instance to be started in low priority for all your frames that will be rendered from there on
Checking low priority in vray rollout help as well, we tend to do both just to be safe

Stan

[eyepiz]

a short update for anyone else using windows 10 and more than 64 cores... I'm very happy to report that the windows 10 anniversary update fixes the start menu and runtime broker issues. Still can't change affinity but learning to deal with that.

Nice! that drove me insane trying to figure it out...so is the searchindex.exe not throwing this error anymore?

[peterguthrie]

havent checked specifically, but runtimebroker doesnt go crazy anymore and I can use the windows button again. phew!