File: README.agent

package info (click to toggle)
likwid 4.2.0-2
  • links: PTS, VCS
  • area: main
  • in suites: stretch
  • size: 11,408 kB
  • ctags: 12,654
  • sloc: ansic: 81,942; perl: 17,226; python: 1,132; makefile: 784; cpp: 239; f90: 162; sh: 85; asm: 40
file content (66 lines) | stat: -rw-r--r-- 3,142 bytes parent folder | download | duplicates (2)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
The likwid-agent application is a daemon that reads hardware performance
counters in a periodic fashion. Which counters can be measured is determined by
the system's CPU architecture. Each architecture has its own set of events and
corresponding counter registers. For the measurement the likwid library is used
and interfaced through the Lua interface. The measured values can be exported in
multiple ways like RRD, syslog or gmetric from the Ganglia Monitoring System.


The configuration file needs to be given at startup and has the following
format:
GROUPPATH <PATH_TO_GROUPS> # default is set during installation
EVENTSET <SPACE_SEPARATED_LIST_OF_GROUPS>
DURATION <TIME_IN_SECONDS_TO_MEASURE_EACH_GROUP>
ACCESSMODE <0/1> # 0 is direct access, 1 forward access to the accessDaemon
LOGPATH <PATH_TO_STORE_LOGFILES> # each montitoring group creates a logfile there named likwid.<GROUP>.log
LOGSTYLE <log/update> # log appends new lines, update clears file previously
GMETRIC <True/False> # send measured values to Gangla
GMETRICPATH <PATH_TO_THE_GMETRIC_EXECUTABLE>
GMETRICCONFIG <EXTRA_CONFIG_OPTIONS_TO_GMETRIC>
RRD <True/False> # write measured values to RRD files, one RRD per group
RRDPATH <PATH_TO_STORE_RRD_FILES>
SYSLOG <True/False> # write measured values to syslog
SYSLOGPRIO <prio> # Use priority level <prio> for syslog, default is local0.notice



The group files cannot lie directly in GROUPPATH, you need to create a folder
with the short name of the architecture like sandybridge or ivybridge. This
enables to use the same group path distributed over a set of systems with different
CPU architecture. The format of a group file is the following:
SHORT <SHORT_NAME_OF_THE GROUP>

EVENTSET // Starts event/counter definitions
FIXC0 INSTR_RETIRED_ANY // Measure event INSTR_RETIRED_ANY in counter FIXC0
FIXC1 CPU_CLK_UNHALTED_CORE
FIXC2 CPU_CLK_UNHALTED_REF
PMC0  L1D_REPLACEMENT
PMC1  L2_TRANS_L1D_WB
PMC2  L2_LINES_IN_ALL
PMC3  L2_LINES_OUT_DEMAND_DIRTY

METRICS // Starts section of derived metrics and output items
ONCE Runtime (RDTSC) [s] time # Output runtime only once
MIN CPI FIXC1/FIXC0 # Output the minimum of the formula FIXC1/FIXC0 named CPI
AVG CPI FIXC1/FIXC0 # Output the average of the same formula
MAX L2 load bandwidth [MBytes/s]  1.0E-06*PMC0*64.0/time # Calculate bandwidth and output only the maximum
MIN L2 load data volume [GBytes]  1.0E-09*PMC0*64.0
SUM L2 bandwidth [MBytes/s] 1.0E-06*(PMC0+PMC1)*64.0/time # Sum up all the values of all CPUs
SUM L2 data volume [GBytes] 1.0E-09*(PMC0+PMC1)*64.0
SUM L3 bandwidth [MBytes/s] 1.0E-06*(PMC2+PMC3)*64.0/time
SUM L3 data volume [GBytes] 1.0E-09*(PMC2+PMC3)*64.0

LONG
<LONG DESCRIPTION OF THE GROUP>

Possible functions are:
ONCE: Output only once (CPU core 0), no aggregation is done
MIN: Output the minimum of all cores
MAX: Output the maximum of all cores
AVG: Output the average of all cores
SUM: Output the sum of all cores' values
If no function is set, the values of all HW threads is written to output and
T<ID> is written in front of the name.

The output metric names can be equal, the function is glued to the output name for later separation.