Understanding CPU Microarchitecture for Performance (JChampionsConf)

@alblue 22 ©2022 Alex Blewitt DMI x4 ** Platform Topologies

8S Configuration SKL SKL LBG LBG LBG DMI LBG SKL SKL SKL SKL SKL SKL 3x16 PCIe* 4S Configurations SKL SKL SKL SKL 2S Configurations SKL SKL (4S-2UPI & 4S-3UPI shown) (2S-2UPI & 2S-3UPI shown) Intel® UPI LBG 3x16 PCIe* 1x100G Intel® OP Fabric 3x16 PCIe* 1x100G Intel® OP Fabric LBG LBG LBG DMI 3x16 PCIe* This slide under embargo until 9:15 AM PDT July 11, 2017 Intel® Xeon® Scalable Processor supports configurations ranging from 2S-2UPI to 8S Non Uniform Memory Architecture (NUMA) https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf

@alblue 22 ©2022 Alex Blewitt DMI x4 ** Platform Topologies

@alblue 22 ©2022 Alex Blewitt https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf 12

@alblue 22 ©2022 Alex Blewitt 18 https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf

@alblue 22 ©2022 Alex Blewitt https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf

@alblue 22 ©2022 Alex Blewitt 10 https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf Cascade/Skylake 10-core die

@alblue 22 ©2022 Alex Blewitt 18 https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf Cascade/Skylake 18-core die

@alblue 22 ©2022 Alex Blewitt https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track-A_1000.pdf Cascade 56 core ‘die’

@alblue 22 ©2022 Alex Blewitt Alder Lake/Saphire Rapids • Alder

Lake is the successor to Ice Lake, built on Willow/Golden Cove • Server will be available in Saphire Rapids • Desktop/laptop cores exist under Alder Lake name • Increases in cache sizes, reduced L1 access latency, higher µop • Moving towards non-heterogenous System on Chip designs for laptops • Mix of E ffi ciency and Power cores in same system • E ffi ciency use less power but are slower

@alblue 22 ©2022 Alex Blewitt L3$ (LLC) fi le 280

Integer 224 Floating Point L1 Data (L1D$) 48 KiB 12-way L1 Instruction (L1I$) 32 KiB 8-way L2$ 512 KiB 8-way Inclusive L3$ (LLC) 1.5 MiB 12-way Non-inclusive Information for Ice Lake (Sunny Cove) RAM RAM RAM RAM 1🕐 5🕐 5🕐 40🕐 12🕐 50🕐 150🕐 300🕐 Clock Cycles

@alblue 22 ©2022 Alex Blewitt Magnetic Tape Magnetic disk Beer

cache hierarchy https://www.slideshare.net/michael_heinrichs/quantum-physics-of-java https://netopyr.com/2014/11/28/reactions-to-the-beer-cache-hierarchy/ Network SSD Main memory L3/LLC L2 L1 Register 🤤 🍺 🍻 🗄 🏪 🚙 ✈ 🍺💰 Beer in mouth Beer in hand Beer in ice bucket Beer in fridge Beer in local store Beer in remote store Beer in another country 🧑🍳Beer being brewed 🌱Beer being planted

@alblue 22 ©2022 Alex Blewitt 💻 lstopo --no-io Machine (16GB)

Package P#0 L4 (128MB) L3 (6144KB) L2 (256KB) L1d (32KB) L1i (32KB) Core P#0 PU P#0 PU P#4 L2 (256KB) L1d (32KB) L1i (32KB) Core P#1 PU P#1 PU P#5 L2 (256KB) L1d (32KB) L1i (32KB) Core P#2 PU P#2 PU P#6 L2 (256KB) L1d (32KB) L1i (32KB) Core P#3 PU P#3 PU P#7 Shared memory between CPU and GPU HyperThreads Single socket system Cache levels Four core processor

@alblue 22 ©2022 Alex Blewitt https://goodies.dotnetos.org/ fi les/dotnetos-poster-ram.pdf

@alblue 22 ©2022 Alex Blewitt L3$ (LLC) fi le 280

Integer 224 Floating Point L1 Data (L1D$) 48 KiB 12-way L1 Instruction (L1I$) 32 KiB 8-way L2$ 512KiB 8-way Inclusive L3$ (LLC) 1.5 MiB 12-way Non-inclusive Data TLB 4K: 128 8-way 2M/4M: 64 8-way 1G: 4 4-way Instruction TLB 4K: 128 8-way 2M/4M: 16/T assoc 1G: 4 4-way STLB 4K: 2048 12-way 2M/4M: 1024 12-way 1G: 1024 4-way RAM RAM RAM RAM Virtual Physical PCID 00008000(1234) 5e38450c(1234) 10 00008000(1234) 48656c6f(1234) 20 ff ff ffff ff fb(8080) 2345 ffff ff b(8080) 0 1🕐 5🕐 5🕐 40🕐 12🕐 50🕐 150🕐 300🕐 Clock Cycles Information for Ice Lake (Sunny Cove) grep /proc/cpuinfo for pcid ↑

@alblue 22 ©2022 Alex Blewitt Memory Pages 8000 ffaa ffbb

f000 0000 CR3 0000 ffff 7fff CR3 0000 ffff 7fff 8000 f000 Two layer page table structure shown x86_64 has 4 level paging (48 bits, 256TiB virtual, 64TiB real) Ice Lake processors support 5 level paging (57 bits, 128Pb virtual, 4PiB real) 0x000080001234 0x000080001234 Pages can be 4k, 2M or 1G

@alblue 22 ©2022 Alex Blewitt 👎 Hugetblfs 👎 • Requires

kernel con fi guration to reserve memory ahead of time • Boot parameter hugepages=N puts aside memory for huge page use • Boot parameter hugepagesz={2M,1G} speci fi es huge page size • Requires a hugetblfs mount to be provided • Requires root (or suitably permissioned app) to use hugepages

@alblue 22 ©2022 Alex Blewitt 👍 Transparent Huge Pages 👎

• Does not require boot time con fi guration or special permissions • khugepaged assembles contiguous physical memory for large pages • Default page size is still 4k, but processes can madvise() use of large pages • Allows speci fi c apps to opt-in on demand • Bene fi ts of smaller TLB with less wasted memory # echo madvise > /sys/kernel/mm/transparent_hugepages/enabled # echo defer > /sys/kernel/mm/transparent_hugepage/defrag Defer instead of blocking large page request Enable opt-in through use of madvise 💡

@alblue 22 ©2022 Alex Blewitt Cache lines, loads and stores

• Unit of granularity of a cache entry is 64/128 bytes (512/1024 bits) • Even if you only read/write 1 byte you're writing 64/128 bytes • Cache lines can generally be in di ff erent states: ➡ M – exclusively owned by that core, and modi fi ed (dirty) ➡ E – exclusively owned by that core, but not modi fi ed ➡ S – shared read-only with other cores ➡ I – invalid, cache line not used Various extensions to MESI exist … see https://en.wikipedia.org/wiki/MESI_protocol for more

@alblue 22 ©2022 Alex Blewitt False sharing • Two cores

trying to write to bytes in the same cache-line will thrash • First thread will try to acquire exclusive ownership of cache line • Second thread (on di ff erent core) will try to do the same • Performance will su ff er when cache line repeatedly moved • Avoid by padding to at least cacheline size * 2 (128/256 bytes) for writes Thread 1 data[0] = 'A' Thread 2 data[7] = 'C'

@alblue 22 ©2022 Alex Blewitt Memory performance strategies • Ensure

data fi ts in L1/L2/L3 cache where possible • Stream or stride through data in a single pass if possible • Consider pivoting data (array-of-structs or structs-of-arrays) • Add padding for multi-threaded contended writes • Prefer thread-local or cpu-local accumulators with fi nal merge step • Compress data where practical (compressed pointers) 💡

@alblue 22 ©2022 Alex Blewitt Pinning memory/threads • Pinning memory

or threads to a particular core can improve performance • Reduces intra-core memory ownership tra ff i c • Less likely to have cache invalidations • isolcpu allows reservation of CPUs for non-kernel use with cpusets • taskset allows binding of a process to speci fi c cores • numactl allows cores/memory to be clamped for a process • libnuma has additional a ff i nity settings for programmatic use

@alblue 22 ©2022 Alex Blewitt Branch Prediction ⤵🧐 • Correct

95% of the time • Queues up instructions assuming the branch has been taken • Learns patterns in code based on existing behaviour • Iterating through predictable (sorted) data may be more e ffi cient • Throws away inaccurate work if incorrect • May cause observable side channel behaviour e.g. cache invalidation 👻 • Unrolled loops and inlining avoid branches, so improves performance cmp eax,42; jne

@alblue 22 ©2022 Alex Blewitt Branch Target Predictor 🎯 •

Predicts where the target is going if taken • Hard coded addresses/o ff sets always predictable • Jump to location of register may be more di ffi cult • Often seen when jumping through object oriented code • Inlining is the master optimisation because it avoids unknowable branches • Dan Luu has a good write up https://danluu.com/branch-prediction/ jmp [eax]

@alblue 22 ©2022 Alex Blewitt Pipeline stalls • CPUs are

deeply pipelined (15-20 stages) • When PC (program counter/instruction pointer) changes causes a fl ush • Gaps occur after jump due to pipeline re fi lling • Good branch predictor guesses correctly to avoid this • HyperThreading can take advantage of 'empty' slots

@alblue 22 ©2022 Alex Blewitt Core Allocate Rename Retire load

buffer 128 entry store buffer 72 entry register fi les 280 + 224 2 3 4 7 8 9 0 1 5 6 Scheduler Integer Unit Floating Unit ALU LEA Shift Branch ALU FMA Shift Divide ALU LEA Multiply Divide ALU FMA Shift Shu ffl e ALU LEA Multiply ALU FMA Shu ffl e ALU LEA Shift Branch Execution units added in Ice Lake Port 0 and 1 can be fused for a 512 bit operation Port 5 is a 512 bit wide operation All others handle 256 bits Port 8 and 9 added in Ice Lake Address generation reorder 352 entry Frontend L1 Data 48 KiB 12-way L1 Instruction 32 KiB 8-way

@alblue 22 ©2022 Alex Blewitt Core Allocate Rename Retire load

@alblue 22 ©2022 Alex Blewitt Micro-architecture diagrams https://drive.google.com/drive/folders/1W4CIRKtNML74BKjSbXerRsIzAUk3ppSG https://twitter.com/Cardyak

@alblue 22 ©2022 Alex Blewitt perf • Linux perf (compiled

from linux/tools/perf, or from linux-tools/linux-perf) • Running in Docker may require compilation from source if kernel mismatch • Commands available • record – record execution performance for process/pid • report – generate a report from prior recording • annotate – annotate a report from a prior recording • stat – record performance counters for process/pid https://perf.wiki.kernel.org https://github.com/alblue/scripts/blob/master/perf-Docker fi le

@alblue 22 ©2022 Alex Blewitt JMH and perf • The

Java Microbenchmarking Harness supports perf natively on Linux • Also able to run on Windows as well using a di ff erent mechanism • perf - runs and collects perf statistics • perfnorm - and normalises them • perfasm - shows assembly output of perf https://github.com/openjdk/jmh

perf • The Async Pro fi ler for Java supports perf output for native JIT • When running in containers need to make sure the perfmap is exported • perf-map-agent - Java agent to generate JIT maps for use with perf • Can be used to generate mixed-mode fl ame graphs (JIT, Java etc.) https://github.com/jvm-pro fi ling-tools

will sample the process(es) and generate stack traces • Events may be skewed from their location • Improve accuracy with :p, :pp or :ppp su ff i x to event • Can capture branches, last branch records or use processor tracing • perf record -b program • perf record --call-graph lbr -j any_call,any_ret program • perf record -e intel_pt//u program https://lwn.net/Articles/680985/ https://lwn.net/Articles/680996/

base64 <(echo hello) d29ybGQK Performance counter stats for 'base64 /dev/fd/63': 0.341382 task-clock (msec) # 0.649 CPUs utilized 0 context-switches # 0.000 K/sec 0 cpu-migrations # 0.000 K/sec 65 page-faults # 0.190 M/sec 1,218,176 cycles # 3.568 GHz 811,468 stalled-cycles-frontend # 66.61% frontend cycles idle 855,999 instructions # 0.70 insn per cycle # 0.95 stalled cycles per insn 169,032 branches # 495.140 M/sec 8,883 branch-misses # 5.26% of all branches 0.000526160 seconds time elapsed https://perf.wiki.kernel.org IPC > 4 👍 < 1 👎

have a few dedicated and programmable counters • Instruction cycles, branches, branch misses … • Counters can be multiplexed (read X for 1µs, read Y for 1µs) • Programmable counters can be set to speci fi c measurements • iTLB-load-misses, LLC-load-misses, uops_dispatched_port.port_5 ... • Undocumented performance counters can be speci fi ed with events • cpu/event=0x3c,umask=0x0,any=1/

MONITORING EVENTS Additionally, the metric uses the UOPS_ISSUED.ANY, which is common in recent Intel microarchitec- tures, as the denominator. The UOPS_ISSUED.ANY event counts the total number of Uops that the RAT issues to RS. The VectorMixRate metric gives the percentage of injected blend uops out of all uops issued. Usually a VectorMixRate over 5% is worth investigating. VectorMixRate[%] = 100 * UOPS_ISSUED.VECTOR_WIDTH_MISMATCH / UOPS_ISSUED.ANY Note the actual penalty may vary as it stems from the additional data-dependency on the destination register the injected blend operations add. B.2 PERFORMANCE MONITORING AND MICROARCHITECTURE This section provides information of performance monitoring hardware and terminology related to the Silvermont, Airmont and Goldmont microarchitectures. The features described here may be specific to individual microarchitecture, as indicated in Table B-1. Figure B-3. TMAM Hierarchy Supported by Skylake Microarchitecture WŝƉĞůŝŶĞ^ůŽƚƐ ZĞƚŝƌŝŶŐ ĂĚ^ƉĞĐƵůĂƚŝŽŶ &ƌŽŶƚŶĚŽƵŶĚ ĂĐŬŶĚŽƵŶĚ EŽƚ^ƚĂůůĞĚ ^ƚĂůůĞĚ ĂƐĞ ƌĂŶĐŚ D ŝƐƉƌĞĚŝĐƚ &ĞƚĐŚ >ĂƚĞŶĐǇ D Ğŵ ŽƌǇŽƵŶĚ ŽƌĞŽƵŶĚ &ĞƚĐŚ ĂŶĚǁ ŝĚƚŚ D ĂĐŚŝŶĞ ůĞĂƌ D ^Ͳ ZKD ǆƚ͘ D Ğŵ ŽƌǇ ŽƵŶĚ >ϯŽƵŶĚ >ϮŽƵŶĚ >ϭŽƵŶĚ ^ƚŽƌĞƐŽƵŶĚ ŝǀŝĚĞƌ ǆĞĐƵƚŝŽŶ ƉŽƌƚƐ hƚŝůŝǌĂƚŝŽŶ >^ D/d ƌĂŶĐŚ ZĞƐƚĞĞƌƐ /ĐĂĐŚĞDŝƐƐ /d>DŝƐƐ KƚŚĞƌ &WͲƌŝƚŚ ^ ^^ǁŝƚĐŚĞƐ D^^ǁŝƚĐŚĞƐ ^ĐĂůĂƌ sĞĐƚŽƌ ϯнƉŽƌƚƐ ϭŽƌϮƉŽƌƚƐ ϬƉŽƌƚƐ DĞŵĂŶĚǁŝĚƚŚ DĞŵ>ĂƚĞŶĐǇ yϴϳ ^ƚŽƌĞDŝƐƐ ^d>,ŝƚ ^d>DŝƐƐ >Ϯ,ŝƚ >ϮDŝƐƐ &ĂůƐĞƐŚĂƌŝŶŐ d>^ƚŽƌĞ ^ƚŽƌĞĨǁĚďůŬ ϰ<ĂůŝĂƐŝŶŐ ŽŶƚĞƐƚĞĚĂĐĐĞƐƐ ĂƚĂƐŚĂƌŝŶŐ >ϯůĂƚĞŶĐǇ USING PERFORMANCE MONITORING EVENTS The single entry point of division at a pipeline’s issue-stage (allocation-stage) makes the four categories additive to the total possible slots. The classification at slots granularity (sub-cycle) makes the break- down very accurate and robust for superscalar cores, which is a necessity at the top-level. Figure B-2. TMAM’s Top Level Drill Down Flowchart hŽƉ ůůŽĐĂƚĞ͍ hŽƉǀĞƌ ZĞƚŝƌĞƐ͍ ĂĐŬŶĚ ^ƚĂůůƐ͍ &ƌŽŶƚŶĚ ŽƵŶĚ ĂĐŬŶĚ ŽƵŶĚ ZĞƚŝƌŝŶŐ ĂĚ ^ƉĞĐƵůĂƚŝŽŶ zĞƐ zĞƐ EŽ zĞƐ EŽ EŽ https://software.intel.com/en-us/download/intel-64-and-ia-32-architectures-optimization-reference-manual Ahmed Yasin

stat -a --topdown sleep 1 nmi_watchdog enabled with topdown. May give wrong results. Disable with echo 0 > /proc/sys/kernel/nmi_watchdog Performance counter stats for 'system wide': retiring bad speculation frontend bound backend bound S0-C0 2 15.3% 2.8% 32.1% 49.9% S0-C1 2 23.3% 4.0% 27.3% 45.4% S0-C2 2 15.2% 2.9% 29.8% 52.1% S0-C3 2 16.7% 0.0% 31.8% 51.5% S0-C4 2 35.7% 10.7% 26.2% 27.4% S0-C5 2 14.9% 2.5% 34.1% 48.5% 1.000889285 seconds time elapsed

Kleen has written toplev.py which allows top-down analysis • Initial download caches processor information from download.01.org • Uses perf to record stats, but with custom event fi lters • If workload is repeatable, can use --no-multiplex to repeat results • Run with -l1, see if issues are present, run with -l2 ... https://github.com/andikleen/pmu-tools/wiki/toplev-manual

of=/tmp/rand bs=4096 count=4096 $ ./toplev.py --single-thread --no-multiplex -l1 -- base64 /tmp/rand > /dev/null # 3.6-full on Intel(R) Xeon(R) CPU E5-1650 v2 @ 3.50GHz BE Backend_Bound % Slots 24.07 <==   $ ./toplev.py --single-thread --no-multiplex -l2 -- base64 /tmp/rand > /dev/null BE Backend_Bound % Slots 23.82 BE/Core Backend_Bound.Core_Bound % Slots 16.08 <==   $ ./toplev.py --single-thread --no-multiplex -l3 -- base64 /tmp/rand > /dev/null BE Backend_Bound % Slots 23.96 BE/Core Backend_Bound.Core_Bound % Slots 16.35 BE/Core Backend_Bound.Core_Bound.Ports_Utilization % Clocks 24.51 <==

Heat maps for instruction memory accesses of the HHVM binary, without and with BOLT. Heat is a log scale. Executed instructions are distributed across icache space After sorting basic blocks guided by pro fi ling data, the icache space is defragmented https://github.com/facebookincubator/BOLT https://github.com/llvm/llvm-project/commit/4c106cfdf7cf7eec861ad3983a3dd9a9e8f3a8ae Now merged into Clang!

is its own networked mesh cluster • Branch speculation and memory/TLB misses are costly • Use branch free and lock free algorithms when possible • Analyse perf counters with top down architectural analysis • Use (auto)vectorisation and use XMM/YMM/ZMM when sensible

@alblue 22 ©2022 Alex Blewitt References https://danluu.com/branch-prediction/ https://arxiv.org/abs/1807.06735 → https://github.com/facebookincubator/BOLT

https://arxiv.org/abs/1902.08318 → https://github.com/lemire/simdjson/ https://github.com/andikleen/pmu-tools/wiki/toplev-manual https://github.com/google/llvm-propeller/ https://lwn.net/Articles/680985/ && https://lwn.net/Articles/680996/ https://perf.wiki.kernel.org https://simplecore-ger.intel.com/swdevcon-uk/wp-content/uploads/sites/5/2017/10/UK-Dev-Con_Toby-Smith-Track- A_1000.pdf https://www.nextplatform.com/2021/04/19/deep-dive-into-intels-ice-lake-xeon-sp-architecture/ https://software.intel.com/en-us/download/intel-64-and-ia-32-architectures-optimization-reference-manual https://www.researchgate.net/publication/269302126_A_Top- Down_method_for_performance_analysis_and_counters_architecture https://drive.google.com/drive/folders/1W4CIRKtNML74BKjSbXerRsIzAUk3ppSG → https://twitter.com/Cardyak https://goodies.dotnetos.org/ fi les/dotnetos-poster-ram.pdf → https://twitter.com/dotnetosorg