Make cpu kind information available #80

carstenbauer · 2024-06-14T09:03:35Z

Use case: Distinguishing efficiency and performance cores.

Close #57

carstenbauer · 2024-06-14T09:04:55Z

On Mac mini M1 (4 efficiency and 4 performance cores)

julia> using Hwloc

julia> num_cpukinds()
2

julia> num_virtual_cores_cpukinds()
2-element Vector{Int64}:
 4
 4

julia> Hwloc.get_cpukind_info()
2-element Vector{Union{Nothing, @NamedTuple{masks::Vector{UInt64}, efficiency_rank::Int32, infos::Vector{Hwloc.HwlocInfo}}}}:
 (masks = UInt64[0x000000000000000f], efficiency_rank = 0, infos = Hwloc.HwlocInfo[Hwloc.HwlocInfo("DarwinCompatible", "apple,icestorm;ARM,v8")])
 (masks = UInt64[0x00000000000000f0], efficiency_rank = 1, infos = Hwloc.HwlocInfo[Hwloc.HwlocInfo("DarwinCompatible", "apple,firestorm;ARM,v8")])

Integration into topology() via the cpukind kwarg:

julia> topology(; cpukind=true)

Machine (3.49 GB)
    Package L#0 P#0 (3.49 GB)
        NUMANode (3.49 GB)
        L2 (4.0 MB)
            L1 (64.0 kB) + Core L#0 P#0
                PU L#0 P#0 (1, DarwinCompatible=apple,icestorm;ARM,v8)
            L1 (64.0 kB) + Core L#1 P#1
                PU L#1 P#1 (1, DarwinCompatible=apple,icestorm;ARM,v8)
            L1 (64.0 kB) + Core L#2 P#2
                PU L#2 P#2 (1, DarwinCompatible=apple,icestorm;ARM,v8)
            L1 (64.0 kB) + Core L#3 P#3
                PU L#3 P#3 (1, DarwinCompatible=apple,icestorm;ARM,v8)
        L2 (12.0 MB)
            L1 (128.0 kB) + Core L#4 P#4
                PU L#4 P#4 (2, DarwinCompatible=apple,firestorm;ARM,v8)
            L1 (128.0 kB) + Core L#5 P#5
                PU L#5 P#5 (2, DarwinCompatible=apple,firestorm;ARM,v8)
            L1 (128.0 kB) + Core L#6 P#6
                PU L#6 P#6 (2, DarwinCompatible=apple,firestorm;ARM,v8)
            L1 (128.0 kB) + Core L#7 P#7
                PU L#7 P#7 (2, DarwinCompatible=apple,firestorm;ARM,v8)
    CoProc(OpenCL) "opencl0d0"

carstenbauer · 2024-06-15T08:30:02Z

debug windows failure

carstenbauer · 2024-06-15T10:02:44Z

I think this is ready for review/merge. (I apologize for the formatting changes in advance 😄.)

I've tested this on a cluster (single kind of CPU cores) and a M1 mac Mini (two kinds of CPU cores).

carstenbauer · 2024-06-17T07:30:18Z

@giordano Can you perhaps run topology(; cpukind=true) on Fugaku? It has a particularly strange mapping of cores to "OS indexes" and I'd like to see what it gives here.

giordano · 2024-06-17T08:08:41Z

I don't have access to Fugaku anymore 😢

JBlaschke · 2024-06-17T09:52:22Z

There seems to be a problem on Perlmutter:

julia> topology(; cpukind=true)

Machine (503.14 GB)
    Package L#0 P#0 (251.18 GB)
        Group (62.21 GB)
            NUMANode (62.21 GB)
            L3 (32.0 MB)
                L2 (512.0 kB) + L1 (32.0 kB) + Core L#0 P#0 ERROR: type Nothing has no field masks
Stacktrace:
  [1] getproperty(x::Nothing, f::Symbol)
    @ Base ./Base.jl:37
  [2] _osindex2cpukind(i::Int64)
    @ Hwloc ~/.julia/dev/Hwloc/src/highlevel_api.jl:469
  [3]
    @ Hwloc ~/.julia/dev/Hwloc/src/highlevel_api.jl:77
  [4]
    @ Hwloc ~/.julia/dev/Hwloc/src/highlevel_api.jl:159
  [5]  (repeats 2 times)
    @ Hwloc ~/.julia/dev/Hwloc/src/highlevel_api.jl:154
  [6]  (repeats 4 times)
    @ Hwloc ~/.julia/dev/Hwloc/src/highlevel_api.jl:159
  [7] print_topology(obj::Hwloc.Object; kwargs::@Kwargs{cpukind::Bool})
    @ Hwloc ~/.julia/dev/Hwloc/src/highlevel_api.jl:175
  [8] print_topology
    @ ~/.julia/dev/Hwloc/src/highlevel_api.jl:175 [inlined]
  [9] topology
    @ ~/.julia/dev/Hwloc/src/highlevel_api.jl:198 [inlined]
 [10] top-level scope
    @ REPL[7]:1
Some type information was truncated. Use `show(err)` to see complete types.

Which seems to stem from the fact that the cpukind info is empty:

julia> num_cpukinds()
1

julia>

julia> num_virtual_cores_cpukinds()

julia> Hwloc.get_cpukind_info()
1-element Vector{Union{Nothing, @NamedTuple{masks::Vector{UInt64}, efficiency_rank::Int32, infos::Vector{Hwloc.HwlocInfo}}}}:
 nothing

Any thoughts?

carstenbauer · 2024-06-17T13:20:17Z

There seems to be a problem on Perlmutter

@JBlaschke Fixed (tested on a Perlmutter login node). Apparently hwloc can't obtain the CPU kind information on Perlmutter and my handling of this case was flawed.

carstenbauer · 2024-06-17T13:39:31Z

(The CI failure seems to be a unrelated hickup)

robertschade · 2024-06-17T13:45:11Z

On an Intel 12900K (8 performance cores, 8 efficiency cores),

julia> num_cpukinds()
2
julia> num_virtual_cores_cpukinds()
2-element Vector{Int64}:
  8
 16
julia> Hwloc.get_cpukind_info()
2-element Vector{Union{Nothing, @NamedTuple{masks::Vector{UInt64}, efficiency_rank::Int32, infos::Vector{Hwloc.HwlocInfo}}}}:
 (masks = UInt64[0x0000000000ff0000], efficiency_rank = 0, infos = Hwloc.HwlocInfo[Hwloc.HwlocInfo("FrequencyMaxMHz", "3900"), Hwloc.HwlocInfo("FrequencyBaseMHz", "2400"), Hwloc.HwlocInfo("CoreType", "IntelAtom")])
 (masks = UInt64[0x000000000000ffff], efficiency_rank = 1, infos = Hwloc.HwlocInfo[Hwloc.HwlocInfo("FrequencyMaxMHz", "5100"), Hwloc.HwlocInfo("FrequencyBaseMHz", "3200"), Hwloc.HwlocInfo("CoreType", "IntelCore")])
julia> topology(; cpukind=true)

Machine (30.64 GB)
    Package L#0 P#0 (30.64 GB)
        NUMANode (30.64 GB)
        L3 (30.0 MB)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#0 P#0 
                PU L#0 P#0 (2, FrequencyMaxMHz=5100; ...)
                PU L#1 P#1 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#1 P#4 
                PU L#2 P#2 (2, FrequencyMaxMHz=5100; ...)
                PU L#3 P#3 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#2 P#8 
                PU L#4 P#4 (2, FrequencyMaxMHz=5100; ...)
                PU L#5 P#5 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#3 P#12 
                PU L#6 P#6 (2, FrequencyMaxMHz=5100; ...)
                PU L#7 P#7 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#4 P#16 
                PU L#8 P#8 (2, FrequencyMaxMHz=5100; ...)
                PU L#9 P#9 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#5 P#20 
                PU L#10 P#10 (2, FrequencyMaxMHz=5100; ...)
                PU L#11 P#11 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#6 P#24 
                PU L#12 P#12 (2, FrequencyMaxMHz=5100; ...)
                PU L#13 P#13 (2, FrequencyMaxMHz=5100; ...)
            L2 (1.25 MB) + L1 (48.0 kB) + Core L#7 P#28 
                PU L#14 P#14 (2, FrequencyMaxMHz=5100; ...)
                PU L#15 P#15 (2, FrequencyMaxMHz=5100; ...)
            L2 (2.0 MB)
                L1 (32.0 kB) + Core L#8 P#32 
                    PU L#16 P#16 (1, FrequencyMaxMHz=3900; ...)
                L1 (32.0 kB) + Core L#9 P#33 
                    PU L#17 P#17 (1, FrequencyMaxMHz=3900; ...)
                L1 (32.0 kB) + Core L#10 P#34 
                    PU L#18 P#18 (1, FrequencyMaxMHz=3900; ...)
                L1 (32.0 kB) + Core L#11 P#35 
                    PU L#19 P#19 (1, FrequencyMaxMHz=3900; ...)
            L2 (2.0 MB)
                L1 (32.0 kB) + Core L#12 P#36 
                    PU L#20 P#20 (1, FrequencyMaxMHz=3900; ...)
                L1 (32.0 kB) + Core L#13 P#37 
                    PU L#21 P#21 (1, FrequencyMaxMHz=3900; ...)
                L1 (32.0 kB) + Core L#14 P#38 
                    PU L#22 P#22 (1, FrequencyMaxMHz=3900; ...)
                L1 (32.0 kB) + Core L#15 P#39 
                    PU L#23 P#23 (1, FrequencyMaxMHz=3900; ...)
    HostBridge 
        PCI 00:02.0 (VGA)
            GPU "renderD128"
            GPU "card0"
        PCIBridge 
            PCI 01:00.0 (NVMExp)
                Block(Disk) "nvme0n1"
        PCIBridge 
            PCI 02:00.0 (NVMExp)
                Block(Disk) "nvme1n1"
        PCIBridge 
            PCI 05:00.0 (Ethernet)
                Net "enp5s0"

carstenbauer · 2024-06-17T14:01:30Z

That looks nice / as expected. Thanks for testing @robertschade!

JBlaschke · 2024-06-17T14:40:54Z

There seems to be a problem on Perlmutter

@JBlaschke Fixed (tested on a Perlmutter login node). Apparently hwloc can't obtain the CPU kind information on Perlmutter and my handling of this case was flawed.

Thanks! Do you know why this information isn't available on PM? We've got normal AMD Milan CPUs there...

carstenbauer · 2024-06-17T18:27:11Z

Do you know why this information isn't available on PM? We've got normal AMD Milan CPUs there...

No idea, hwloc is my abstraction and it doesn't provide it (lstopo --cpukinds also prints nothing).

carstenbauer · 2024-06-17T18:28:03Z

FYI: I plan to merge this tomorrow.

JBlaschke · 2024-06-18T09:45:35Z

No idea, hwloc is my abstraction and it doesn't provide it (lstopo --cpukinds also prints nothing).

Ok so that's probably our fault then -- I'll follow up internatlly.

FYI: I plan to merge this tomorrow.

Brilliant! Thanks for your work on this :)

first attempt to make cpukinds information available

a935a61

carstenbauer changed the title ~~Make cpukinds information available~~ Make cpu kind information available Jun 14, 2024

carstenbauer added 14 commits June 14, 2024 11:11

improve docstring

91c970c

fix

81a5e00

fix

a66be55

tmp

9cc179f

support multiple ulongs

b8f79fd

rename num_virtual_cores_cpukinds

7279100

minor rename

dc04b1a

cleanup

dd06e07

first basic tests

4c696e9

extract cpu kind info + prepare printing

ad437fb

fix tests

7bcea3b

printing

f3a9171

minor

827aaa6

cpukind and cpukindinfo kwargs

10b5dbb

carstenbauer added 4 commits June 15, 2024 11:17

try fix windows ci

afaaab1

try fix windows ci x86

244ec12

drop cpukindinfo / merge it into cpukind

0ad804b

Mention cpu kind features in README

ae1eef9

carstenbauer requested review from eschnett and JBlaschke June 15, 2024 10:03

carstenbauer added 2 commits June 15, 2024 12:07

minor

24a0ba4

support multiple masks (on systems with > 64 virtual cores)

6c71634

fix Perlmutter issue + improve case of unavailable cpu kind info

0e232c0

JBlaschke approved these changes Jun 17, 2024

View reviewed changes

carstenbauer merged commit 2ee0444 into master Jun 18, 2024
39 of 40 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Make cpu kind information available #80

Make cpu kind information available #80

carstenbauer commented Jun 14, 2024 •

edited

Loading

carstenbauer commented Jun 14, 2024 •

edited

Loading

carstenbauer commented Jun 15, 2024 •

edited

Loading

carstenbauer commented Jun 15, 2024

carstenbauer commented Jun 17, 2024

giordano commented Jun 17, 2024

JBlaschke commented Jun 17, 2024

carstenbauer commented Jun 17, 2024 •

edited

Loading

carstenbauer commented Jun 17, 2024

robertschade commented Jun 17, 2024

carstenbauer commented Jun 17, 2024

JBlaschke commented Jun 17, 2024

carstenbauer commented Jun 17, 2024

carstenbauer commented Jun 17, 2024

JBlaschke commented Jun 18, 2024

Make cpu kind information available #80

Make cpu kind information available #80

Conversation

carstenbauer commented Jun 14, 2024 • edited Loading

carstenbauer commented Jun 14, 2024 • edited Loading

carstenbauer commented Jun 15, 2024 • edited Loading

carstenbauer commented Jun 15, 2024

carstenbauer commented Jun 17, 2024

giordano commented Jun 17, 2024

JBlaschke commented Jun 17, 2024

carstenbauer commented Jun 17, 2024 • edited Loading

carstenbauer commented Jun 17, 2024

robertschade commented Jun 17, 2024

carstenbauer commented Jun 17, 2024

JBlaschke commented Jun 17, 2024

carstenbauer commented Jun 17, 2024

carstenbauer commented Jun 17, 2024

JBlaschke commented Jun 18, 2024

carstenbauer commented Jun 14, 2024 •

edited

Loading

carstenbauer commented Jun 14, 2024 •

edited

Loading

carstenbauer commented Jun 15, 2024 •

edited

Loading

carstenbauer commented Jun 17, 2024 •

edited

Loading