6 perf-c2c - Shared Data C2C/HITM Analyzer.
11 'perf c2c record' [<options>] <command>
12 'perf c2c record' [<options>] -- [<record command options>] <command>
13 'perf c2c report' [<options>]
17 C2C stands for Cache To Cache.
19 The perf c2c tool provides means for Shared Data C2C/HITM analysis. It allows
20 you to track down the cacheline contentions.
22 On x86, the tool is based on load latency and precise store facility events
23 provided by Intel CPUs. On PowerPC, the tool uses random instruction sampling
24 with thresholding feature.
27 - memory address of the access
28 - type of the access (load and store details)
29 - latency (in cycles) of the load access
31 The c2c tool provide means to record this data and report back access details
32 for cachelines with highest contention - highest number of HITM accesses.
34 The basic workflow with this tool follows the standard record/report phase.
35 User uses the record command to record events data and report command to
43 Select the PMU event. Use 'perf c2c record -e list'
44 to list available events.
48 Be more verbose (show counter open errors, etc).
52 Configure mem-loads latency. (x86 only)
56 Configure all used events to run in kernel space.
60 Configure all used events to run in user space.
70 Be more verbose (show counter open errors, etc).
74 Specify the input file to process.
78 Show extra node info in report (see NODE INFO section)
82 Specify sorting fields for single cacheline display.
83 Following fields are available: tid,pid,iaddr,dso
88 Setup callchains parameters.
89 Please refer to perf-report man page for details.
92 Force the stdio output (see STDIO OUTPUT)
95 Display only statistic tables and force stdio mode.
98 Display full length of symbols.
101 Do not display Source:Line column.
104 Show all captured HITM lines, with no regard to HITM % 0.0005 limit.
108 Don't do ownership validation.
112 Switch to HITM type (rmt, lcl) to display and sort on. Total HITMs as default.
115 Show callgraph with stitched LBRs, which may have more complete
116 callgraph. The perf.data file must have been obtained using
117 perf c2c record --call-graph lbr.
118 Disabled by default. In common cases with call stack overflows,
119 it can recreate better call stacks than the default lbr call stack
120 output. But this approach is not full proof. There can be cases
121 where it creates incorrect call stacks from incorrect matches.
122 The known limitations include exception handing such as
123 setjmp/longjmp will have calls/returns not match.
127 The perf c2c record command setup options related to HITM cacheline analysis
128 and calls standard perf record command.
130 Following perf record options are configured by default:
131 (check perf record man page for details)
133 -W,-d,--phys-data,--sample-cpu
135 Unless specified otherwise with '-e' option, following events are monitored by
138 cpu/mem-loads,ldlat=30/P
141 and following on PowerPC:
146 User can pass any 'perf record' option behind '--' mark, like (to enable
147 callchains and system wide monitoring):
149 $ perf c2c record -- -g -a
151 Please check RECORD OPTIONS section for specific c2c record options.
155 The perf c2c report command displays shared data analysis. It comes in two
156 display modes: stdio and tui (default).
158 The report command workflow is following:
159 - sort all the data based on the cacheline address
160 - store access details for each cacheline
161 - sort all cachelines based on user settings
164 In general perf report output consist of 2 basic views:
165 1) most expensive cachelines list
166 2) offsets details for each cacheline
168 For each cacheline in the 1) list we display following data:
169 (Both stdio and TUI modes follow the same fields output)
172 - zero based index to identify the cacheline
175 - cacheline address (hex number)
178 - cacheline percentage of all Remote/Local HITM accesses
180 LLC Load Hitm - Total, LclHitm, RmtHitm
181 - count of Total/Local/Remote load HITMs
184 - sum of all cachelines accesses
187 - sum of all load accesses
190 - sum of all store accesses
192 Store Reference - L1Hit, L1Miss
193 L1Hit - store accesses that hit L1
194 L1Miss - store accesses that missed L1
196 Core Load Hit - FB, L1, L2
197 - count of load hits in FB (Fill Buffer), L1 and L2 cache
199 LLC Load Hit - LlcHit, LclHitm
200 - count of LLC load accesses, includes LLC hits and LLC HITMs
202 RMT Load Hit - RmtHit, RmtHitm
203 - count of remote load accesses, includes remote hits and remote HITMs
206 - count of local and remote DRAM accesses
208 For each offset in the 2) list we display following data:
211 - % of Remote/Local HITM accesses for given offset within cacheline
213 Store Refs - L1 Hit, L1 Miss
214 - % of store accesses that hit/missed L1 for given offset within cacheline
216 Data address - Offset
220 - pid of the process responsible for the accesses
223 - tid of the process responsible for the accesses
226 - code address responsible for the accesses
228 cycles - rmt hitm, lcl hitm, load
229 - sum of cycles for given accesses - Remote/Local HITM and generic load
232 - number of cpus that participated on the access
235 - code symbol related to the 'Code address' value
238 - shared object name related to the 'Code address' value
241 - source information related to the 'Code address' value
244 - nodes participating on the access (see NODE INFO section)
248 The 'Node' field displays nodes that accesses given cacheline
249 offset. Its output comes in 3 flavors:
250 - node IDs separated by ','
251 - node IDs with stats for each ID, in following format:
252 Node{cpus %hitms %stores}
253 - node IDs with list of affected CPUs in following format:
256 User can switch between above flavors with -N option or
257 use 'n' key to interactively switch in TUI mode.
261 User can specify how to sort offsets for cacheline.
263 Following fields are available and governs the final
264 output fields set for caheline offsets output:
266 tid - coalesced by process TIDs
267 pid - coalesced by process PIDs
268 iaddr - coalesced by code address, following fields are displayed:
269 Code address, Code symbol, Shared Object, Source line
270 dso - coalesced by shared object
272 By default the coalescing is setup with 'pid,iaddr'.
276 The stdio output displays data on standard output.
278 Following tables are displayed:
279 Trace Event Information
280 - overall statistics of memory accesses
282 Global Shared Cache Line Event Information
283 - overall statistics on shared cachelines
285 Shared Data Cache Line Table
286 - list of most expensive cachelines
288 Shared Cache Line Distribution Pareto
289 - list of all accessed offsets for each cacheline
293 The TUI output provides interactive interface to navigate
294 through cachelines list and to display offset details.
296 For details please refer to the help window by pressing '?' key.
300 Although Don Zickus, Dick Fowles and Joe Mario worked together
301 to get this implemented, we got lots of early help from Arnaldo
302 Carvalho de Melo, Stephane Eranian, Jiri Olsa and Andi Kleen.
306 Check Joe's blog on c2c tool for detailed use case explanation:
307 https://joemario.github.io/blog/2016/09/01/c2c-blog/
311 linkperf:perf-record[1], linkperf:perf-mem[1]