性能測試

Android 8.0 includes binder and hwbinder performance tests for throughput and latency. While many scenarios exist for detecting perceptible performance problems, running such scenarios can be time consuming and results are often unavailable until after a system is integrated. Using the provided performance tests makes it easier to test during development, detect serious problems earlier, and improve user experience.

Performance tests include the following four categories:

• binder throughput (available in system/libhwbinder/vts/performance/Benchmark_binder.cpp)
• binder latency (available in frameworks/native/libs/binder/tests/schd-dbg.cpp)
• hwbinder throughput (available in system/libhwbinder/vts/performance/Benchmark.cpp)
• hwbinder latency (available in system/libhwbinder/vts/performance/Latency.cpp)

About binder and hwbinder

Binder and hwbinder are Android inter-process communication (IPC) infrastructures that share the same Linux driver but have the following qualitative differences:

Binder and hwbinder processes

A systrace visualizer displays transactions as follows:

In the above example:

• The four (4) schd-dbg processes are client processes.
• The four (4) binder processes are server processes (name starts with Binder and ends with a sequence number).
• A client process is always paired with a server process, which is dedicated to its client.
• All the client-server process pairs are scheduled independently by kernel concurrently.

In CPU 1, the OS kernel executes the client to issue the request. It then uses the same CPU whenever possible to wake up a server process, handle the request, and context switch back after the request is complete.

Throughput vs. latency

In a perfect transaction, where the client and server process switch seamlessly, throughput and latency tests do not produce substantially different messages. However, when the OS kernel is handling an interrupt request (IRQ) from hardware, waiting for locks, or simply choosing not to handle a message immediately, a latency bubble can form.

The throughput test generates a large number of transactions with different payload sizes, providing a good estimation for the regular transaction time (in best case scenarios) and the maximum throughput the binder can achieve.

In contrast, the latency test performs no actions on the payload to minimize the regular transaction time. We can use transaction time to estimate the binder overhead, make statistics for the worst case, and calculate the ratio of transactions whose latency meets a specified deadline.

Handling priority inversions

A priority inversion occurs when a thread with higher priority is logically waiting for a thread with lower priority. Real-time (RT) applications have a priority inversion problem:

When using Linux Completely Fair Scheduler (CFS) scheduling, a thread always has a chance to run even when other threads have a higher priority. As a result, applications with CFS scheduling handle priority inversion as expected behavior and not as a problem. In cases where the Android framework needs RT scheduling to guarantee the privilege of high priority threads however, priority inversion must be resolved.

Example priority inversion during a binder transaction (RT thread is logically blocked by other CFS threads when waiting for a binder thread to service):

To avoid blockages, you can use priority inheritance to temporarily escalate the Binder thread to a RT thread when it services a request from a RT client. Keep in mind that RT scheduling has limited resources and should be used carefully. In a system with n CPUs, the maximum number of current RT threads is also n; additional RT threads might need to wait (and thus miss their deadlines) if all CPUs are taken by other RT threads.

To resolve all possible priority inversions, you could use priority inheritance for both binder and hwbinder. However, as binder is widely used across the system, enabling priority inheritance for binder transactions might spam the system with more RT threads than it can service.

Running throughput tests

The throughput test is run against binder/hwbinder transaction throughput. In a system that is not overloaded, latency bubbles are rare and their impact can be eliminated as long as the number of iterations is high enough.

• The binder throughput test is in system/libhwbinder/vts/performance/Benchmark_binder.cpp.
• The hwbinder throughput test is in system/libhwbinder/vts/performance/Benchmark.cpp.

Test results

Example throughput test results for transactions using different payload sizes:

Benchmark                      Time          CPU           Iterations
---------------------------------------------------------------------
BM_sendVec_binderize/4         70302 ns      32820 ns      21054
BM_sendVec_binderize/8         69974 ns      32700 ns      21296
BM_sendVec_binderize/16        70079 ns      32750 ns      21365
BM_sendVec_binderize/32        69907 ns      32686 ns      21310
BM_sendVec_binderize/64        70338 ns      32810 ns      21398
BM_sendVec_binderize/128       70012 ns      32768 ns      21377
BM_sendVec_binderize/256       69836 ns      32740 ns      21329
BM_sendVec_binderize/512       69986 ns      32830 ns      21296
BM_sendVec_binderize/1024      69714 ns      32757 ns      21319
BM_sendVec_binderize/2k        75002 ns      34520 ns      20305
BM_sendVec_binderize/4k        81955 ns      39116 ns      17895
BM_sendVec_binderize/8k        95316 ns      45710 ns      15350
BM_sendVec_binderize/16k      112751 ns      54417 ns      12679
BM_sendVec_binderize/32k      146642 ns      71339 ns       9901
BM_sendVec_binderize/64k      214796 ns     104665 ns       6495

• Time indicates the round trip delay measured in real time.
• CPU indicates the accumulated time when CPUs are scheduled for the test.
• Iterations indicates the number of times the test function executed.

For example, for an 8-byte payload:

BM_sendVec_binderize/8         69974 ns      32700 ns      21296

… the maximum throughput the binder can achieve is calculated as:

MAX throughput with 8-byte payload = (8 * 21296)/69974 ~= 2.423 b/ns ~= 2.268 Gb/s

Test options

To get results in .json, run the test with the --benchmark_format=json argument:

libhwbinder_benchmark --benchmark_format=json
{
  "context": {
    "date": "2017-05-17 08:32:47",
    "num_cpus": 4,
    "mhz_per_cpu": 19,
    "cpu_scaling_enabled": true,
    "library_build_type": "release"
  },
  "benchmarks": [
    {
      "name": "BM_sendVec_binderize/4",
      "iterations": 32342,
      "real_time": 47809,
      "cpu_time": 21906,
      "time_unit": "ns"
    },
   ….
}

Running latency tests

The latency test measures the time it takes for the client to begin initializing the transaction, switch to the server process for handling, and receive the result. The test also looks for known bad scheduler behaviors that can negatively impact transaction latency, such as a scheduler that does not support priority inheritance or honor the sync flag.

• The binder latency test is in frameworks/native/libs/binder/tests/schd-dbg.cpp.
• The hwbinder latency test is in system/libhwbinder/vts/performance/Latency.cpp.

Test results

Results (in .json) show statistics for average/best/worst latency and the number of deadlines missed.

Test options

Latency tests take the following options:

The following sections detail each option, describe usage, and provide example results.

Specifying iterations

Example with a large number of iterations and verbose output disabled:

libhwbinder_latency -i 5000 -pair 3
{
"cfg":{"pair":3,"iterations":5000,"deadline_us":2500},
"P0":{"SYNC":"GOOD","S":9352,"I":10000,"R":0.9352,
  "other_ms":{ "avg":0.2 , "wst":2.8 , "bst":0.053, "miss":2, "meetR":0.9996},
  "fifo_ms": { "avg":0.16, "wst":1.5 , "bst":0.067, "miss":0, "meetR":1}
},
"P1":{"SYNC":"GOOD","S":9334,"I":10000,"R":0.9334,
  "other_ms":{ "avg":0.19, "wst":2.9 , "bst":0.055, "miss":2, "meetR":0.9996},
  "fifo_ms": { "avg":0.16, "wst":3.1 , "bst":0.066, "miss":1, "meetR":0.9998}
},
"P2":{"SYNC":"GOOD","S":9369,"I":10000,"R":0.9369,
  "other_ms":{ "avg":0.19, "wst":4.8 , "bst":0.055, "miss":6, "meetR":0.9988},
  "fifo_ms": { "avg":0.15, "wst":1.8 , "bst":0.067, "miss":0, "meetR":1}
},
"inheritance": "PASS"
}

These test results show the following:

"pair":3

Creates one client and server pair.

"iterations": 5000

Includes 5000 iterations.

"deadline_us":2500

Deadline is 2500us (2.5ms); most transactions are expected to meet this value.

"I": 10000

A single test iteration includes two (2) transactions:

• One transaction by normal priority (CFS other)
• One transaction by real time priority (RT-fifo)

5000 iterations equals a total of 10000 transactions.

"S": 9352

9352 of the transactions are synced in the same CPU.

"R": 0.9352

Indicates the ratio at which the client and server are synced together in the same CPU.

"other_ms":{ "avg":0.2 , "wst":2.8 , "bst":0.053, "miss":2, "meetR":0.9996}

The average (avg), worst (wst), and the best (bst) case for all transactions issued by a normal priority caller. Two transactions miss the deadline, making the meet ratio (meetR) 0.9996.

"fifo_ms": { "avg":0.16, "wst":1.5 , "bst":0.067, "miss":0, "meetR":1}

Similar to other_ms, but for transactions issued by client with rt_fifo priority. It's likely (but not required) that the fifo_ms has a better result than other_ms, with lower avg and wst values and a higher meetR (the difference can be even more significant with load in the background).

Note: Background load may impact the throughput result and the other_ms tuple in the latency test. Only the fifo_ms may show similar results as long as the background load has a lower priority than RT-fifo.

Specifying pair values

Each client process is paired with a server process dedicated for the client, and each pair may be scheduled independently to any CPU. However, the CPU migration should not happen during a transaction as long as the SYNC flag is honor.

Ensure the system is not overloaded! While high latency in an overloaded system is expected, test results for an overloaded system do not provide useful information. To test a system with higher pressure, use -pair #cpu-1 (or -pair #cpu with caution). Testing using -pair n with n > #cpu overloads the system and generates useless information.

Specifying deadline values

After extensive user scenario testing (running the latency test on a qualified product), we determined that 2.5ms is the deadline to meet. For new applications with higher requirements (such as 1000 photos/second), this deadline value will change.

Specifying verbose output

Using the -v option displays verbose output. Example:

libhwbinder_latency -i 1 -v

--------------------------------------------------
service      pid: 8674 tid: 8674 cpu: 1
SCHED_OTHER 0
--------------------------------------------------
main         pid: 8673 tid: 8673 cpu: 1

--------------------------------------------------
client       pid: 8677 tid: 8677 cpu: 0
SCHED_OTHER 0

--------------------------------------------------
fifo-caller  pid: 8677 tid: 8678 cpu: 0
SCHED_FIFO  99

--------------------------------------------------
hwbinder     pid: 8674 tid: 8676 cpu: 0
???         99
--------------------------------------------------
other-caller pid: 8677 tid: 8677 cpu: 0
SCHED_OTHER 0

--------------------------------------------------
hwbinder     pid: 8674 tid: 8676 cpu: 0
SCHED_OTHER 0

• The service thread is created with a SCHED_OTHER priority and run in CPU:1 with pid 8674.
• The first transaction is then started by a fifo-caller. To service this transaction, the hwbinder upgrades the priority of server (pid: 8674 tid: 8676) to be 99 and also marks it with a transient scheduling class (printed as ???). The scheduler then puts the server process in CPU:0 to run and syncs it with the same CPU with its client.
• The second transaction caller has a SCHED_OTHER priority. The server downgrades itself and services the caller with SCHED_OTHER priority.

Using trace for debugging

You can specify the -trace option to debug latency issues. When used, the latency test stops the tracelog recording at the moment when bad latency is detected. Example:

atrace --async_start -b 8000 -c sched idle workq binder_driver sync freq
libhwbinder_latency -deadline_us 50000 -trace -i 50000 -pair 3
deadline triggered: halt ∓ stop trace
log:/sys/kernel/debug/tracing/trace

The following components can impact latency:

• Android build mode. Eng mode is usually slower than userdebug mode.
• Framework. How does the framework service use ioctl to config to the binder?
• Binder driver. Does the driver support fine-grained locking? Does it contain all performance turning patches?
• Kernel version. The better real time capability the kernel has, the better the results.
• Kernel config. Does the kernel config contain DEBUG configs such as DEBUG_PREEMPT and DEBUG_SPIN_LOCK?
• Kernel scheduler. Does the kernel have an Energy-Aware scheduler (EAS) or Heterogeneous Multi-Processing (HMP) scheduler? Do any kernel drivers (cpu-freq driver, cpu-idle driver, cpu-hotplug, etc.) impact the scheduler?