Description - phoronix/bullet

This is a benchmark of the Bullet Physics Engine.

This benchmark is a single-threaded benchmark with seven workloads that run in seconds. In the order they are run:

• raytests
• 3000 fall
• 1000 stack
• 1000 convex
• 136 ragdolls
• prim trimesh
• convex trimesh

These tests are all run from a single application and then summarized later. The tests supposedly report times in “seconds”, but this isn’t quite correct as the entire application runs in ~5 seconds total.

Results for 3000 fall: 4.812264
Results for 1000 stack: 5.579135
Results for 136 ragdolls: 3.058886
Results for 1000 convex: 5.174400
Results for prim-trimesh: 1.045790
Results for convex-trimesh: 1.277420
Results for raytests: 2.928500

All tests were run pinned to core 1.

The output of a run can be seen at the following link.

With these quick running tests in a single application, one can’t separate out the tests.

Metrics (Intel) - phoronix/bullet

sh - pid 11861
	On_CPU   0.125
	On_Core  0.998
	IPC      1.866
	Retire   0.478	(47.8%)
	FrontEnd 0.067	(6.7%)
	Spec     0.083	(8.3%)
	Backend  0.372	(37.2%)
	Elapsed   5.11
	Procs    3
	Maxrss   37K
	Minflt   16638
	Majflt   0
	Inblock  0
	Oublock  64
	Msgsnd   0
	Msgrcv   0
	Nsignals 0
	Nvcsw    18	(62.1%)
	Nivcsw   11
	Utime    5.089383
	Stime    0.012000
	Start    652669.13
	Finish   652674.24

Close to 100% of time of one core, with backend stalls the largest issue.

Metrics (AMD) - phoronix/bullet

sh - pid 21387
	On_CPU   0.062
	On_Core  1.000
	IPC      2.031
	FrontCyc 0.160	(16.0%)
	BackCyc  0.151	(15.1%)
	Elapsed   4.87
	Procs    3
	Maxrss   37K
	Minflt   16642
	Majflt   0
	Inblock  0
	Oublock  64
	Msgsnd   0
	Msgrcv   0
	Nsignals 0
	Nvcsw    18	(3.6%)
	Nivcsw   484
	Utime    4.854848
	Stime    0.013637
	Start    658760.91
	Finish   658765.78

Process Tree - phoronix/bullet
Process Tree
The process tree is simple

    11861) sh
      11862) bullet
        11863) AppBenchmarks

100% of time is spent running on one core.

Some noise in the IPC but difficult to be certain as the benchmark runs so quickly.

Backend stalls are the largest issue. In this plot the frontend stalls come during post-processing operations.

Topdown (Intel)

retire         0.256
ms_uops                0.111
speculation    0.053
branch_misses          19.85%
machine_clears         80.15%
frontend       0.288
idq_uops_delivered_0   0.117
idq_uops_delivered_1   0.140
idq_uops_delivered_2   0.160
idq_uops_delivered_3   0.184
backend        0.402
resource_stalls.sb     0.080
stalls_ldm_pending     0.554

Speculation misses appear to be primarily machine clears. The number of macro uops is also larger than average.