1perf-c2c(1) 2=========== 3 4NAME 5---- 6perf-c2c - Shared Data C2C/HITM Analyzer. 7 8SYNOPSIS 9-------- 10[verse] 11'perf c2c record' [<options>] <command> 12'perf c2c record' [<options>] \-- [<record command options>] <command> 13'perf c2c report' [<options>] 14 15DESCRIPTION 16----------- 17C2C stands for Cache To Cache. 18 19The perf c2c tool provides means for Shared Data C2C/HITM analysis. It allows 20you to track down the cacheline contentions. 21 22On Intel, the tool is based on load latency and precise store facility events 23provided by Intel CPUs. On PowerPC, the tool uses random instruction sampling 24with thresholding feature. On AMD, the tool uses IBS op pmu (due to hardware 25limitations, perf c2c is not supported on Zen3 cpus). 26 27These events provide: 28 - memory address of the access 29 - type of the access (load and store details) 30 - latency (in cycles) of the load access 31 32The c2c tool provide means to record this data and report back access details 33for cachelines with highest contention - highest number of HITM accesses. 34 35The basic workflow with this tool follows the standard record/report phase. 36User uses the record command to record events data and report command to 37display it. 38 39 40RECORD OPTIONS 41-------------- 42-e:: 43--event=:: 44 Select the PMU event. Use 'perf c2c record -e list' 45 to list available events. 46 47-v:: 48--verbose:: 49 Be more verbose (show counter open errors, etc). 50 51-l:: 52--ldlat:: 53 Configure mem-loads latency. Supported on Intel and Arm64 processors 54 only. Ignored on other archs. 55 56-k:: 57--all-kernel:: 58 Configure all used events to run in kernel space. 59 60-u:: 61--all-user:: 62 Configure all used events to run in user space. 63 64REPORT OPTIONS 65-------------- 66-k:: 67--vmlinux=<file>:: 68 vmlinux pathname 69 70-v:: 71--verbose:: 72 Be more verbose (show counter open errors, etc). 73 74-i:: 75--input:: 76 Specify the input file to process. 77 78-N:: 79--node-info:: 80 Show extra node info in report (see NODE INFO section) 81 82-c:: 83--coalesce:: 84 Specify sorting fields for single cacheline display. 85 Following fields are available: tid,pid,iaddr,dso 86 (see COALESCE) 87 88-g:: 89--call-graph:: 90 Setup callchains parameters. 91 Please refer to perf-report man page for details. 92 93--stdio:: 94 Force the stdio output (see STDIO OUTPUT) 95 96--stats:: 97 Display only statistic tables and force stdio mode. 98 99--full-symbols:: 100 Display full length of symbols. 101 102--no-source:: 103 Do not display Source:Line column. 104 105--show-all:: 106 Show all captured HITM lines, with no regard to HITM % 0.0005 limit. 107 108-f:: 109--force:: 110 Don't do ownership validation. 111 112-d:: 113--display:: 114 Switch to HITM type (rmt, lcl) or peer snooping type (peer) to display 115 and sort on. Total HITMs (tot) as default, except Arm64 uses peer mode 116 as default. 117 118--stitch-lbr:: 119 Show callgraph with stitched LBRs, which may have more complete 120 callgraph. The perf.data file must have been obtained using 121 perf c2c record --call-graph lbr. 122 Disabled by default. In common cases with call stack overflows, 123 it can recreate better call stacks than the default lbr call stack 124 output. But this approach is not full proof. There can be cases 125 where it creates incorrect call stacks from incorrect matches. 126 The known limitations include exception handing such as 127 setjmp/longjmp will have calls/returns not match. 128 129C2C RECORD 130---------- 131The perf c2c record command setup options related to HITM cacheline analysis 132and calls standard perf record command. 133 134Following perf record options are configured by default: 135(check perf record man page for details) 136 137 -W,-d,--phys-data,--sample-cpu 138 139Unless specified otherwise with '-e' option, following events are monitored by 140default on Intel: 141 142 cpu/mem-loads,ldlat=30/P 143 cpu/mem-stores/P 144 145following on AMD: 146 147 ibs_op// 148 149and following on PowerPC: 150 151 cpu/mem-loads/ 152 cpu/mem-stores/ 153 154User can pass any 'perf record' option behind '--' mark, like (to enable 155callchains and system wide monitoring): 156 157 $ perf c2c record -- -g -a 158 159Please check RECORD OPTIONS section for specific c2c record options. 160 161C2C REPORT 162---------- 163The perf c2c report command displays shared data analysis. It comes in two 164display modes: stdio and tui (default). 165 166The report command workflow is following: 167 - sort all the data based on the cacheline address 168 - store access details for each cacheline 169 - sort all cachelines based on user settings 170 - display data 171 172In general perf report output consist of 2 basic views: 173 1) most expensive cachelines list 174 2) offsets details for each cacheline 175 176For each cacheline in the 1) list we display following data: 177(Both stdio and TUI modes follow the same fields output) 178 179 Index 180 - zero based index to identify the cacheline 181 182 Cacheline 183 - cacheline address (hex number) 184 185 Rmt/Lcl Hitm (Display with HITM types) 186 - cacheline percentage of all Remote/Local HITM accesses 187 188 Peer Snoop (Display with peer type) 189 - cacheline percentage of all peer accesses 190 191 LLC Load Hitm - Total, LclHitm, RmtHitm (For display with HITM types) 192 - count of Total/Local/Remote load HITMs 193 194 Load Peer - Total, Local, Remote (For display with peer type) 195 - count of Total/Local/Remote load from peer cache or DRAM 196 197 Total records 198 - sum of all cachelines accesses 199 200 Total loads 201 - sum of all load accesses 202 203 Total stores 204 - sum of all store accesses 205 206 Store Reference - L1Hit, L1Miss, N/A 207 L1Hit - store accesses that hit L1 208 L1Miss - store accesses that missed L1 209 N/A - store accesses with memory level is not available 210 211 Core Load Hit - FB, L1, L2 212 - count of load hits in FB (Fill Buffer), L1 and L2 cache 213 214 LLC Load Hit - LlcHit, LclHitm 215 - count of LLC load accesses, includes LLC hits and LLC HITMs 216 217 RMT Load Hit - RmtHit, RmtHitm 218 - count of remote load accesses, includes remote hits and remote HITMs; 219 on Arm neoverse cores, RmtHit is used to account remote accesses, 220 includes remote DRAM or any upward cache level in remote node 221 222 Load Dram - Lcl, Rmt 223 - count of local and remote DRAM accesses 224 225For each offset in the 2) list we display following data: 226 227 HITM - Rmt, Lcl (Display with HITM types) 228 - % of Remote/Local HITM accesses for given offset within cacheline 229 230 Peer Snoop - Rmt, Lcl (Display with peer type) 231 - % of Remote/Local peer accesses for given offset within cacheline 232 233 Store Refs - L1 Hit, L1 Miss, N/A 234 - % of store accesses that hit L1, missed L1 and N/A (no available) memory 235 level for given offset within cacheline 236 237 Data address - Offset 238 - offset address 239 240 Pid 241 - pid of the process responsible for the accesses 242 243 Tid 244 - tid of the process responsible for the accesses 245 246 Code address 247 - code address responsible for the accesses 248 249 cycles - rmt hitm, lcl hitm, load (Display with HITM types) 250 - sum of cycles for given accesses - Remote/Local HITM and generic load 251 252 cycles - rmt peer, lcl peer, load (Display with peer type) 253 - sum of cycles for given accesses - Remote/Local peer load and generic load 254 255 cpu cnt 256 - number of cpus that participated on the access 257 258 Symbol 259 - code symbol related to the 'Code address' value 260 261 Shared Object 262 - shared object name related to the 'Code address' value 263 264 Source:Line 265 - source information related to the 'Code address' value 266 267 Node 268 - nodes participating on the access (see NODE INFO section) 269 270NODE INFO 271--------- 272The 'Node' field displays nodes that accesses given cacheline 273offset. Its output comes in 3 flavors: 274 - node IDs separated by ',' 275 - node IDs with stats for each ID, in following format: 276 Node{cpus %hitms %stores} (Display with HITM types) 277 Node{cpus %peers %stores} (Display with peer type) 278 - node IDs with list of affected CPUs in following format: 279 Node{cpu list} 280 281User can switch between above flavors with -N option or 282use 'n' key to interactively switch in TUI mode. 283 284COALESCE 285-------- 286User can specify how to sort offsets for cacheline. 287 288Following fields are available and governs the final 289output fields set for cacheline offsets output: 290 291 tid - coalesced by process TIDs 292 pid - coalesced by process PIDs 293 iaddr - coalesced by code address, following fields are displayed: 294 Code address, Code symbol, Shared Object, Source line 295 dso - coalesced by shared object 296 297By default the coalescing is setup with 'pid,iaddr'. 298 299STDIO OUTPUT 300------------ 301The stdio output displays data on standard output. 302 303Following tables are displayed: 304 Trace Event Information 305 - overall statistics of memory accesses 306 307 Global Shared Cache Line Event Information 308 - overall statistics on shared cachelines 309 310 Shared Data Cache Line Table 311 - list of most expensive cachelines 312 313 Shared Cache Line Distribution Pareto 314 - list of all accessed offsets for each cacheline 315 316TUI OUTPUT 317---------- 318The TUI output provides interactive interface to navigate 319through cachelines list and to display offset details. 320 321For details please refer to the help window by pressing '?' key. 322 323CREDITS 324------- 325Although Don Zickus, Dick Fowles and Joe Mario worked together 326to get this implemented, we got lots of early help from Arnaldo 327Carvalho de Melo, Stephane Eranian, Jiri Olsa and Andi Kleen. 328 329C2C BLOG 330-------- 331Check Joe's blog on c2c tool for detailed use case explanation: 332 https://joemario.github.io/blog/2016/09/01/c2c-blog/ 333 334SEE ALSO 335-------- 336linkperf:perf-record[1], linkperf:perf-mem[1] 337