MBA is inaccurate in Intel(R) Xeon(R) Gold 5220R CPU

liuyingdong-byte commented 3 years ago

some information for my test environment

OS information and kernel version Debian GNU/Linux 10 (buster) 4.19.0-10-amd64
cpu information Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian Address sizes: 46 bits physical, 48 bits virtual CPU(s): 96 On-line CPU(s) list: 0-95 Thread(s) per core: 2 Core(s) per socket: 24 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 85 Model name: Intel(R) Xeon(R) Gold 5220R CPU @ 2.20GHz Stepping: 7 CPU MHz: 2901.122 CPU max MHz: 4000.0000 CPU min MHz: 1000.0000 BogoMIPS: 4400.00 Virtualization: VT-x L1d cache: 32K L1i cache: 32K L2 cache: 1024K L3 cache: 36608K NUMA node0 CPU(s): 0-23,48-71 NUMA node1 CPU(s): 24-47,72-95 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 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 intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid 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 avx512_vnni md_clear flush_l1d arch_capabilities
memory information Number Of Devices: 12 Size: 32 GB Type: DDR4 Speed: 2933 MT/s Manufacturer: Samsung Configured Memory Speed: 2666 MT/s
testing method

I use a “mlc” tool from intel: https://software.intel.com/content/www/us/en/develop/articles/intelr-memory-latency-checker.html

The test script is as follows：

#!/bin/bash

core_list="
1
1,2
" # will run cores in this list, 1 loop per line.

operation="-R" # MLC operation, can be change to “W6” or “W7“

function start()
{
        for i in $core_list
        do
                for mba in $(seq 10 10 100)
                do
                        pqos -I -R
                        pqos -I -a core:1=$i   # bound COS1 for cores
                        pqos -I -e mba:1=$mba  # bound MBA policy to COS1

                        echo -n "COS1 cpus_list:"
                        cat /sys/fs/resctrl/COS1/cpus_list
                        echo -n "COS1 schemata:"
                        cat /sys/fs/resctrl/COS1/schemata

                        mkdir -p $i

                        pqos -I -r -m all:[$i] > $i/pqos_$mba.txt & # pqos as monitor

                        echo "mlc --loaded_latency -d0 $operation -t300 -T -k$i > $i/$mba.txt" # run 5min
                        mlc --loaded_latency -d0 $operation -t300 -T -k$i > $i/$mba.txt # run 5min
                        pkill -9 pqos # close pqos
                        pqos -I -R # reset all RDT  settings.

                done
        done
}

start # begin test.`

test results

Memory Bandwidth Allocation Percent	“-k1” pqos MBL[MB/s] avg	“-k1” Error Percent	“-k1” mlc --loaded_latency（Bandwidth MB/sec）	“-k1” Error Percent	“-k1,2” pqos MBL[MB/s] avg	“-k1,2” Error Percent	“-k1,2” mlc --loaded_latency（Bandwidth MB/sec）	“-k1,2” Error Percent
10	1885.79	-50.55%	2310.8	-39.40%	3910.12	-43.74%	4623.1	-41.19%
20	2701.7	-7.84%	3468.5	-4.62%	5862.85	-7.76%	6938.7	-5.95%
30	3682.57	2.00%	4716.1	5.16%	7912.79	3.04%	9430.1	4.00%
40	5991.09	-19.57%	8119.1	-22.45%	13446	-23.57%	16114.4	-23.03%
50	6633.05	-5.91%	8864.5	-6.95%	14511	-6.69%	17576.1	-7.35%
60	7270.89	3.25%	9788.4	1.58%	15638.6	4.19%	19370.9	1.40%
70	8109.8	7.51%	10948	5.65%	17447.1	8.38%	21606.6	5.73%
80	8025.33	19.91%	10949	17.43%	17659.2	18.85%	21609.1	17.51%
90	7867.21	30.21%	10952.6	26.58%	18166.7	25.80%	21614.3	26.66%
100	9896.43	20.99%	13755.1	17.02%	21352.9	21.51%	25468.2	22.22%

From the point, when the restriction percentage is 20, 30, 50, 60, and 70, the restriction effect is better, and other errors are larger.

my question

Has someone ever meet the same things? Please help me, thanks.

limitedge commented 3 years ago

Have you tried this experiment in more than one system and you did try it multiple times? what version of python are you running?

liuyingdong-byte commented 3 years ago

I am sorry but It’s not convenient for me to provide this email address because it’s a private work email. What is the test result with this script I provided? How did you test when control memory bandwidth by unit of percentage(MBA in M%, e.g. pqos -I -e 'mba:1=M')？I find much better when control memory bandwidth by using in MBA in N MB/s(e.g. pqos -I -e 'mba_max:1=N'), did you meet similar situation before？

Thank you for your attention to this matter.

intel / intel-cmt-cat