How to discover which caches (L1,L2,L3) are shared by which HW threads (cores) ?

[gilvannco]

Hi,

how would you discover the information described in subject on a Linux OS? I think that parsing /proc/cpuinfo can give much of the information, but not all. For example, if I have a Quad Core processor, then discovering which cores create the pairs that share L2 isn't possible, right?

I assume that in Quad Core processor, HT threads share L1 cache, then two pairs share L2 cache (so there are two distinct L2 cache units), and then all 4 cores share the same L3 cache. Is this correct?

Regards,

Gilian

[Dmitry_Vyukov]

HT threads definitely share L1.

L3 is shared between all cores.

But AFAIK L2 is private for each core. I.e. 4 cores and 4 L2 caches. (I assume that you mean Core i7 processors, since you mention L3). What you are talking about - 2 L2 caches shared by a pair of cores - was featured on Core Quad (Q6600) processors.

The quick way to verify an assumption is to download CPUZ utility and observe how many L2 caches you have. If it's 4 then they are not shared, if it's 2 then they are shared.

Also take a look at:

http://software.intel.com/en-us/articles/intel-64-architecture-processor-topology-enumeration/

The article describes how to determine cache parameters by an output of CPUID instruction.

[TimP]

The linux application 'irqbalance --debug' and the Intel MPI command cpuinfo attempt to supply cache sharing information:

[tim@tim-blfd bin64]$ ./cpuinfo
Genuine Intel Processor (Intel64)
===== Processor composition =====
Processors(CPUs) : 4
Packages(sockets) : 1
Cores per package : 4
Threads per core : 1
===== Processor identification =====
Processor Thread Id. Core Id. Package Id.
0 0 0 0
1 0 1 0
2 0 2 0
3 0 3 0
===== Placement on packages =====
Package Id. Core Id. Processors
0 0,1,2,3 0,1,2,3
===== Cache sharing =====
Cache Size Processors
L1 32 KB no sharing
L2 256 KB no sharing
L3 8 MB (0,1,2,3)

[tim@tim-blfd bin64]$ /usr/sbin/irqbalance --debug
Package 0: cpu mask is 0000000f (workload 0)
Cache domain 3: cpu mask is 00000008 (workload 0)
CPU number 3 (workload 0)
Cache domain 2: cpu mask is 00000004 (workload 0)
CPU number 2 (workload 0)
Cache domain 1: cpu mask is 00000002 (workload 0)
CPU number 1 (workload 0)
Cache domain 0: cpu mask is 00000001 (workload 0)
CPU number 0 (workload 0)
Interrupt 98 (class ethernet) has workload 8
Interrupt 0 (class timer) has workload 2002
Interrupt 82 (class storage) has workload 22
Interrupt 58 (class storage) has workload 5
Interrupt 177 (class storage) has workload 0
Interrupt 74 (class legacy) has workload 0



---------------------------------------------------------------
IRQ delta is 3062
Package 0: cpu mask is 0000000f (workload 10260)
Cache domain 3: cpu mask is 00000008 (workload 25)
CPU number 3 (workload 24)
Interrupt 98 (ethernet/23)
Interrupt 177 (storage/0)
Cache domain 2: cpu mask is 00000004 (workload 96)
CPU number 2 (workload 0)
Interrupt 82 (storage/95)
Cache domain 1: cpu mask is 00000002 (workload 5)
CPU number 1 (workload 0)
Interrupt 58 (storage/3)
Interrupt 74 (legacy/0)
Cache domain 0: cpu mask is 00000001 (workload 10134)
CPU number 0 (workload 10134)

[gilvannco]

Thank you very much for the answers. I've got following output from cpuinfo:

Intel Xeon Processor (Intel64 Harpertown)
===== Processor composition =====
Processors(CPUs) : 8
Packages(sockets) : 2
Cores per package : 4
Threads per core : 1
===== Processor identification =====
Processor Thread Id. Core Id. Package Id.
0 0 0 0
1 0 0 1
2 0 2 0
3 0 2 1
4 0 1 0
5 0 1 1
6 0 3 0
7 0 3 1
===== Placement on packages =====
Package Id. Core Id. Processors
0 0,2,1,3 0,2,4,6
1 0,2,1,3 1,3,5,7
===== Cache sharing =====
Cache Size Processors
L1 32 KB no sharing
L2 6 MB (0,4)(1,5)(2,6)(3,7)

So, it looks like no L3 cache? And four L2 "banks", each shared by two cores?

Regards,

Gilian

[TimP]

Yes, that's the usual layout for Harpertown.