Add top down metrics to perf stat
From: Andi Kleen
Date: Tue Jan 19 2016 - 21:30:29 EST
[v2: Address review feedback.
Metrics are now always printed, but colored when crossing threshold.
--topdown implies --metric-only.
Various smaller fixes, see individual patches]
[v3: Add --single-thread option and support it with HT off.
Clean up old HT workaround.
Various smaller fixes, see individual patches.
Note to reviewers: includes both tools and kernel patches.
The kernel patches are at the end.
This patchkit adds support for TopDown measurements to perf stat
It applies on top of my earlier metrics patchkit, posted
TopDown is intended to replace the frontend cycles idle/
backend cycles idle metrics in standard perf stat output.
These metrics are not reliable in many workloads,
due to out of order effects.
This implements a new --topdown mode in perf stat
(similar to --transaction) that measures the pipe line
bottlenecks using standardized formulas. The measurement
can be all done with 5 counters (one fixed counter)
The result are four metrics:
FrontendBound, BackendBound, BadSpeculation, Retiring
that describe the CPU pipeline behavior on a high level.
FrontendBound and BackendBound
BadSpeculation is a higher
The full top down methology has many hierarchical metrics.
This implementation only supports level 1 which can be
collected without multiplexing. A full implementation
of top down on top of perf is available in pmu-tools toplev.
The current version works on Intel Core CPUs starting
with Sandy Bridge, and Atom CPUs starting with Silvermont.
In principle the generic metrics should be also implementable
on other out of order CPUs.
TopDown level 1 uses a set of abstracted metrics which
are generic to out of order CPU cores (although some
CPUs may not implement all of them):
topdown-total-slots Available slots in the pipeline
topdown-slots-issued Slots issued into the pipeline
topdown-slots-retired Slots successfully retired
topdown-fetch-bubbles Pipeline gaps in the frontend
topdown-recovery-bubbles Pipeline gaps during recovery
These metrics then allow to compute four useful metrics:
FrontendBound, BackendBound, Retiring, BadSpeculation.
The formulas to compute the metrics are generic, they
only change based on the availability on the abstracted
The kernel declares the events supported by the current
CPU and perf stat then computes the formulas based on the
$ perf stat --topdown -I 1000 cmd
1.000735655 frontend bound retiring bad speculation backend bound
1.000735655 S0-C0 2 47.84% 11.69% 8.37% 32.10%
1.000735655 S0-C1 2 45.53% 11.39% 8.52% 34.56%
2.003978563 S0-C0 2 49.47% 12.22% 8.65% 29.66%
2.003978563 S0-C1 2 47.21% 12.98% 8.77% 31.04%
3.004901432 S0-C0 2 49.35% 12.26% 8.68% 29.70%
3.004901432 S0-C1 2 47.23% 12.67% 8.76% 31.35%
4.005766611 S0-C0 2 48.44% 12.14% 8.59% 30.82%
4.005766611 S0-C1 2 46.07% 12.41% 8.67% 32.85%
5.006580592 S0-C0 2 47.91% 12.08% 8.57% 31.44%
5.006580592 S0-C1 2 45.57% 12.27% 8.63% 33.53%
6.007545125 S0-C0 2 47.45% 12.02% 8.57% 31.96%
6.007545125 S0-C1 2 45.13% 12.17% 8.57% 34.14%
7.008539347 S0-C0 2 47.07% 12.03% 8.61% 32.29%
For Level 1 Top Down computes metrics per core instead of per logical CPU
on Core CPUs (On Atom CPUs there is no Hyper Threading and TopDown
is per thread)
In this case perf stat automatically enables --per-core mode and also requires
global mode (-a) and avoiding other filters (no cgroup mode)
When Hyper Threading is off this can be overriden with the --single-thread
option. When Hyper Threading is on it is enforced, the only way to
not require -a here is to off line the logical CPUs of the second
One side effect is that this may require root rights or a
Full tree available in