Is AMD CPU more sensitive to false-sharing than Intel and Power?

I tried to come up with a robust and minimal false sharing example (so that I can play with perf c2c). But it turned out the following example (from the previous discussion in Zulip) only shows the effect of false sharing in AMD CPU and not in Intel or Power. Does it mean AMD CPU is sensitive to this particular type (I found another example that works outside AMD) of false sharing? I’m also very puzzled why there is no effect in Intel and Power.

MWE

Here’s the code:

function sum_bad(xs)
    accs = zeros(eltype(xs), Threads.nthreads())
    Threads.@threads :static for x in xs
        @inbounds accs[Threads.threadid()] += x
    end
    return sum(accs)
end

const CACHELINE = try
    parse(Int, read("/sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size", String))
catch err
    @warn "cannot read cache line size" exception = (err, catch_backtrace())
    64
end

function sum_padded(xs)
    stride = cld(CACHELINE, sizeof(eltype(xs)))
    accs = zeros(eltype(xs), Threads.nthreads() * stride)
    Threads.@threads :static for x in xs
        @inbounds accs[1 + (Threads.threadid() - 1) * stride] += x
    end
    return sum(accs)
end

xs = rand(2^20);
@assert sum_bad(xs) ≈ sum_padded(xs);

using BenchmarkTools
xs = rand(2^30);
@btime sum_bad($xs);
@btime sum_padded($xs);

I also checked with FLoops.jl version which behaves more-or-less like sum_padded

using FLoops

function sum_floop(xs)
    z = zero(eltype(xs))
    @floop for x in xs
        @reduce(acc = z + x)
    end
    return acc
end

@btime sum_floop($xs);

Result

All using julia1.6 -t8. No special care for reducing “OS jitter” (e.g., processor affinity).

AMD

You can see that the first result is much slower (which is what I wanted)

  920.942 ms (44 allocations: 3.84 KiB)  # sum_bad
  153.757 ms (43 allocations: 4.28 KiB)  # sum_padded
  148.622 ms (68 allocations: 3.94 KiB)  # sum_floop

lscpu output:

Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              128
On-line CPU(s) list: 0-127
Thread(s) per core:  2
Core(s) per socket:  32
Socket(s):           2
NUMA node(s):        2
Vendor ID:           AuthenticAMD
CPU family:          23
Model:               49
Model name:          AMD EPYC 7502 32-Core Processor
Stepping:            0
CPU MHz:             1499.967
CPU max MHz:         2500.0000
CPU min MHz:         1500.0000
BogoMIPS:            5000.16
Virtualization:      AMD-V
L1d cache:           32K
L1i cache:           32K
L2 cache:            512K
L3 cache:            16384K
NUMA node0 CPU(s):   0-31,64-95
NUMA node1 CPU(s):   32-63,96-127
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate sme ssbd mba sev ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif umip rdpid overflow_recov succor smca

Intel

  230.459 ms (42 allocations: 3.78 KiB)
  279.687 ms (42 allocations: 4.25 KiB)
  237.340 ms (68 allocations: 3.94 KiB)
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              40
On-line CPU(s) list: 0-39
Thread(s) per core:  2
Core(s) per socket:  10
Socket(s):           2
NUMA node(s):        2
Vendor ID:           GenuineIntel
CPU family:          6
Model:               85
Model name:          Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz
Stepping:            4
CPU MHz:             800.841
CPU max MHz:         3000.0000
CPU min MHz:         800.0000
BogoMIPS:            4400.00
Virtualization:      VT-x
L1d cache:           32K
L1i cache:           32K
L2 cache:            1024K
L3 cache:            14080K
NUMA node0 CPU(s):   0-9,20-29
NUMA node1 CPU(s):   10-19,30-39
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d

Power

  281.393 ms (42 allocations: 7.53 KiB)
  249.961 ms (42 allocations: 8.53 KiB)
  246.897 ms (69 allocations: 7.25 KiB)
Architecture:          ppc64le
Byte Order:            Little Endian
CPU(s):                160
On-line CPU(s) list:   0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29,32,33,36,37,40,41,44,45,48,49,52,53,56,57,60,61,64,65,68,69,72,73,76,77,80,81,84,85,88,89,92,93,96,97,100,101,104,105,108,109,112,113,116,117,120,121,124,125,128,129,132,133,136,137,140,141,144,145,148,149,152,153,156,157
Off-line CPU(s) list:  2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,34,35,38,39,42,43,46,47,50,51,54,55,58,59,62,63,66,67,70,71,74,75,78,79,82,83,86,87,90,91,94,95,98,99,102,103,106,107,110,111,114,115,118,119,122,123,126,127,130,131,134,135,138,139,142,143,146,147,150,151,154,155,158,159
Thread(s) per core:    2
Core(s) per socket:    20
Socket(s):             2
NUMA node(s):          6
Model:                 2.1 (pvr 004e 1201)
Model name:            POWER9, altivec supported
CPU max MHz:           3800.0000
CPU min MHz:           2300.0000
L1d cache:             32K
L1i cache:             32K
L2 cache:              512K
L3 cache:              10240K
NUMA node0 CPU(s):     0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29,32,33,36,37,40,41,44,45,48,49,52,53,56,57,60,61,64,65,68,69,72,73,76,77
NUMA node8 CPU(s):     80,81,84,85,88,89,92,93,96,97,100,101,104,105,108,109,112,113,116,117,120,121,124,125,128,129,132,133,136,137,140,141,144,145,148,149,152,153,156,157
NUMA node252 CPU(s):
NUMA node253 CPU(s):
NUMA node254 CPU(s):
NUMA node255 CPU(s):
3 Likes

I have no knowledge or thoughts to share about the issue, but out of curiosity I ran the code on my intel i5-11600K CPU with 16 GB RAM (so I reduced 2^30 to 2^20), and with julia -t4 my outputs were:

  231.067 μs (22 allocations: 1.97 KiB) # bad
  231.197 μs (22 allocations: 2.19 KiB) # padded
  230.949 μs (31 allocations: 1.72 KiB) # floop

so this at least reproduces that even on a consumer intel CPU the “bad” sum doesn’t seem to come at any performance cost.

1 Like