3.1.2. RT-linux Performance Guide

Read This First

All performance numbers provided in this document are gathered using following Evaluation Modules unless otherwise specified.

Name Description
AM64x EVM AM64x Evaluation Module rev E1 with ARM running at 1GHz, DDRdata rate 1600 MT/S

Table: Evaluation Modules


About This Manual

This document provides performance data for each of the device drivers which are part of the Process SDK Linux package. This document should be used in conjunction with release notes and user guides provided with the Process SDK Linux package for information on specific issues present with drivers included in a particular release.

If You Need Assistance

For further information or to report any problems, contact http://e2e.ti.com/ or http://support.ti.com/

3.1.2.1. System Benchmarks

3.1.2.1.1. LMBench

LMBench is a collection of microbenchmarks of which the memory bandwidth and latency related ones are typically used to estimate processor memory system performance.

Latency: lat_mem_rd-stride128-szN, where N is equal to or smaller than the cache size at given level measures the cache miss penalty. N that is at least double the size of last level cache is the latency to external memory.

Bandwidth: bw_mem_bcopy-N, where N is is equal to or smaller than the cache size at a given level measures the achivable memory bandwidth from software doing a memcpy() type operation. Typical use is for external memory bandwidth calculation. The bandwidth is calculated as byte read and written counts as 1 which should be roughly half of STREAM copy result.

Benchmarks am64xx-evm: perf
af_unix_sock_stream_latency (microsec) 41.81
af_unix_socket_stream_bandwidth (MBs) 655.38
bw_file_rd-io-1mb (MB/s) 813.54
bw_file_rd-o2c-1mb (MB/s) 473.48
bw_mem-bcopy-16mb (MB/s) 860.59
bw_mem-bcopy-1mb (MB/s) 918.61
bw_mem-bcopy-2mb (MB/s) 895.66
bw_mem-bcopy-4mb (MB/s) 890.37
bw_mem-bcopy-8mb (MB/s) 909.19
bw_mem-bzero-16mb (MB/s) 2118.92
bw_mem-bzero-1mb (MB/s) 1514.90 (min 918.61, max 2111.19)
bw_mem-bzero-2mb (MB/s) 1505.85 (min 895.66, max 2116.03)
bw_mem-bzero-4mb (MB/s) 1503.76 (min 890.37, max 2117.15)
bw_mem-bzero-8mb (MB/s) 1489.39 (min 909.19, max 2069.59)
bw_mem-cp-16mb (MB/s) 522.35
bw_mem-cp-1mb (MB/s) 1398.83 (min 526.22, max 2271.44)
bw_mem-cp-2mb (MB/s) 1354.70 (min 529.17, max 2180.23)
bw_mem-cp-4mb (MB/s) 1343.43 (min 528.96, max 2157.89)
bw_mem-cp-8mb (MB/s) 1375.33 (min 605.60, max 2145.06)
bw_mem-fcp-16mb (MB/s) 772.61
bw_mem-fcp-1mb (MB/s) 1456.11 (min 801.03, max 2111.19)
bw_mem-fcp-2mb (MB/s) 1473.70 (min 831.37, max 2116.03)
bw_mem-fcp-4mb (MB/s) 1492.37 (min 867.58, max 2117.15)
bw_mem-fcp-8mb (MB/s) 1492.57 (min 915.54, max 2069.59)
bw_mem-frd-16mb (MB/s) 1368.69
bw_mem-frd-1mb (MB/s) 1060.77 (min 801.03, max 1320.51)
bw_mem-frd-2mb (MB/s) 1087.05 (min 831.37, max 1342.73)
bw_mem-frd-4mb (MB/s) 1114.53 (min 867.58, max 1361.47)
bw_mem-frd-8mb (MB/s) 1148.86 (min 915.54, max 1382.17)
bw_mem-fwr-16mb (MB/s) 2120.33
bw_mem-fwr-1mb (MB/s) 1795.98 (min 1320.51, max 2271.44)
bw_mem-fwr-2mb (MB/s) 1761.48 (min 1342.73, max 2180.23)
bw_mem-fwr-4mb (MB/s) 1759.68 (min 1361.47, max 2157.89)
bw_mem-fwr-8mb (MB/s) 1763.62 (min 1382.17, max 2145.06)
bw_mem-rd-16mb (MB/s) 1381.33
bw_mem-rd-1mb (MB/s) 713.86 (min 661.55, max 766.17)
bw_mem-rd-2mb (MB/s) 774.72 (min 723.33, max 826.10)
bw_mem-rd-4mb (MB/s) 1092.66 (min 817.33, max 1367.99)
bw_mem-rd-8mb (MB/s) 1126.60 (min 880.28, max 1372.92)
bw_mem-rdwr-16mb (MB/s) 860.17
bw_mem-rdwr-1mb (MB/s) 677.64 (min 526.22, max 829.05)
bw_mem-rdwr-2mb (MB/s) 655.67 (min 529.17, max 782.17)
bw_mem-rdwr-4mb (MB/s) 651.25 (min 528.96, max 773.54)
bw_mem-rdwr-8mb (MB/s) 736.55 (min 605.60, max 867.49)
bw_mem-wr-16mb (MB/s) 912.30
bw_mem-wr-1mb (MB/s) 797.61 (min 766.17, max 829.05)
bw_mem-wr-2mb (MB/s) 804.14 (min 782.17, max 826.10)
bw_mem-wr-4mb (MB/s) 795.44 (min 773.54, max 817.33)
bw_mem-wr-8mb (MB/s) 873.89 (min 867.49, max 880.28)
bw_mmap_rd-mo-1mb (MB/s) 1274.81
bw_mmap_rd-o2c-1mb (MB/s) 482.16
bw_pipe (MB/s) 288.15
bw_unix (MB/s) 655.38
lat_connect (us) 78.52
lat_ctx-2-128k (us) 5.91
lat_ctx-2-256k (us) 4.78
lat_ctx-4-128k (us) 8.77
lat_ctx-4-256k (us) 1.70
lat_fs-0k (num_files) 203.00
lat_fs-10k (num_files) 83.00
lat_fs-1k (num_files) 128.00
lat_fs-4k (num_files) 144.00
lat_mem_rd-stride128-sz1000k (ns) 47.13
lat_mem_rd-stride128-sz125k (ns) 7.83
lat_mem_rd-stride128-sz250k (ns) 12.73
lat_mem_rd-stride128-sz31k (ns) 5.12
lat_mem_rd-stride128-sz50 (ns) 3.01
lat_mem_rd-stride128-sz500k (ns) 43.70
lat_mem_rd-stride128-sz62k (ns) 7.38
lat_mmap-1m (us) 76.00
lat_ops-double-add (ns) 0.73
lat_ops-double-mul (ns) 4.02
lat_ops-float-add (ns) 0.73
lat_ops-float-mul (ns) 4.02
lat_ops-int-add (ns) 1.01
lat_ops-int-bit (ns) 0.67
lat_ops-int-div (ns) 6.04
lat_ops-int-mod (ns) 6.40
lat_ops-int-mul (ns) 3.05
lat_ops-int64-add (ns) 1.01
lat_ops-int64-bit (ns) 0.67
lat_ops-int64-div (ns) 9.56
lat_ops-int64-mod (ns) 7.40
lat_pagefault (us) 2.20
lat_pipe (us) 20.85
lat_proc-exec (us) 1506.50
lat_proc-fork (us) 1345.25
lat_proc-proccall (us) 0.01
lat_select (us) 43.79
lat_sem (us) 2.48
lat_sig-catch (us) 7.94
lat_sig-install (us) 0.84
lat_sig-prot (us) 0.55
lat_syscall-fstat (us) 2.10
lat_syscall-null (us) 0.36
lat_syscall-open (us) 395.79
lat_syscall-read (us) 0.93
lat_syscall-stat (us) 5.53
lat_syscall-write (us) 0.61
lat_tcp (us) 0.70
lat_unix (us) 41.81
latency_for_0.50_mb_block_size (nanosec) 43.70
latency_for_1.00_mb_block_size (nanosec) 23.57 (min 0.00, max 47.13)
pipe_bandwidth (MBs) 288.15
pipe_latency (microsec) 20.85
procedure_call (microsec) 0.01
select_on_200_tcp_fds (microsec) 43.79
semaphore_latency (microsec) 2.48
signal_handler_latency (microsec) 0.84
signal_handler_overhead (microsec) 7.94
tcp_ip_connection_cost_to_localhost (microsec) 78.52
tcp_latency_using_localhost (microsec) 0.70

Table: LM Bench Metrics

3.1.2.1.2. Dhrystone

Dhrystone is a core only benchmark that runs from warm L1 caches in all modern processors. It scales linearly with clock speed. For standard ARM cores the DMIPS/MHz score will be identical with the same compiler and flags.

Benchmarks am64xx-evm: perf
cpu_clock (MHz) 1000.00
dhrystone_per_mhz (DMIPS/MHz) 2.99
dhrystone_per_second (DhrystoneP) 5263158.00

Table: Dhrystone Benchmark

3.1.2.1.3. Whetstone

Benchmarks am64xx-evm: perf
whetstone (MIPS) 1000.00

Table: Whetstone Benchmark

3.1.2.1.4. Linpack

Linpack measures peak double precision (64 bit) floating point performance in sloving a dense linear system.

Benchmarks am64xx-evm: perf
linpack (Kflops) 391598.00

Table: Linpack Benchmark

3.1.2.1.5. NBench

Benchmarks am64xx-evm: perf
assignment (Iterations) 9.76
fourier (Iterations) 16312.00
fp_emulation (Iterations) 76.43
huffman (Iterations) 837.28
idea (Iterations) 2450.30
lu_decomposition (Iterations) 391.56
neural_net (Iterations) 5.60
numeric_sort (Iterations) 359.48
string_sort (Iterations) 118.26

Table: NBench Benchmarks

3.1.2.1.6. Stream

STREAM is a microbenchmarks for measuring data memory system performance without any data reuse. It is designed to miss on caches and exercise data prefetcher and apeculative accesseses. it uses double precision floating point (64bit) but in most modern processors the memory access will be the bottleck. The four individual scores are copy, scale as in multiply by constant, add two numbers, and triad for multiply accumulate. For bandwidth a byte read counts as one and a byte written counts as one resulting in a score that is double the bandwidth LMBench will show.

Benchmarks am64xx-evm: perf
add (MB/s) 466.80
copy (MB/s) 519.30
scale (MB/s) 569.50
triad (MB/s) 440.90

Table: Stream

3.1.2.1.7. CoreMarkPro

CoreMark®-Pro is a comprehensive, advanced processor benchmark that works with and enhances the market-proven industry-standard EEMBC CoreMark® benchmark. While CoreMark stresses the CPU pipeline, CoreMark-Pro tests the entire processor, adding comprehensive support for multicore technology, a combination of integer and floating-point workloads, and data sets for utilizing larger memory subsystems.

Benchmarks am64xx-evm: perf
cjpeg-rose7-preset (workloads/) 6.31
core (workloads/) 0.09
coremark-pro () 332.14
linear_alg-mid-100x100-sp (workloads/) 10.46
loops-all-mid-10k-sp (workloads/) 0.49
nnet_test (workloads/) 0.78
parser-125k (workloads/) 5.59
radix2-big-64k (workloads/) 5.41
sha-test (workloads/) 13.19
zip-test (workloads/) 15.38

Table: CoreMarkPro

Benchmarks am64xx-evm: perf
cjpeg-rose7-preset (workloads/) 49.02
core (workloads/) 0.43
coremark-pro () 1011.02
linear_alg-mid-100x100-sp (workloads/) 20.06
loops-all-mid-10k-sp (workloads/) 0.88
nnet_test (workloads/) 1.54
parser-125k (workloads/) 6.45
radix2-big-64k (workloads/) 31.21
sha-test (workloads/) 93.46
zip-test (workloads/) 28.99

Table: CoreMarkPro for Two Cores

3.1.2.1.8. MultiBench

MultiBench™ is a suite of benchmarks that allows processor and system designers to analyze, test, and improve multicore processors. It uses three forms of concurrency: Data decomposition: multiple threads cooperating on achieving a unified goal and demonstrating a processor’s support for fine grain parallelism. Processing multiple data streams: uses common code running over multiple threads and demonstrating how well a processor scales over scalable data inputs. Multiple workload processing: shows the scalability of general-purpose processing, demonstrating concurrency over both code and data. MultiBench combines a wide variety of application-specific workloads with the EEMBC Multi-Instance-Test Harness (MITH), compatible and portable with most any multicore processors and operating systems. MITH uses a thread-based API (POSIX-compliant) to establish a common programming model that communicates with the benchmark through an abstraction layer and provides a flexible interface to allow a wide variety of thread-enabled workloads to be tested.

Benchmarks am64xx-evm: perf
4m-check (workloads/) 330.78
4m-check-reassembly (workloads/) 61.65
4m-check-reassembly-tcp (workloads/) 37.65
4m-check-reassembly-tcp-cmykw2-rotatew2 (workloads/) 15.29
4m-check-reassembly-tcp-x264w2 (workloads/) 0.76
4m-cmykw2 (workloads/) 87.07
4m-cmykw2-rotatew2 (workloads/) 18.15
4m-reassembly (workloads/) 50.45
4m-rotatew2 (workloads/) 21.32
4m-tcp-mixed (workloads/) 86.02
4m-x264w2 (workloads/) 0.24
empty-wld (workloads/) 1.00
idct-4m (workloads/) 13.66
idct-4mw1 (workloads/) 3.85
ippktcheck-4m (workloads/) 329.90
ippktcheck-4mw1 (workloads/) 89.47
ipres-4m (workloads/) 68.52
ipres-4mw1 (workloads/) 68.59
md5-4m (workloads/) 19.95
md5-4mw1 (workloads/) 4.07
rgbcmyk-4m (workloads/) 44.16
rgbcmyk-4mw1 (workloads/) 44.21
rotate-4ms1 (workloads/) 16.91
rotate-4ms1w1 (workloads/) 16.95
rotate-4ms64 (workloads/) 17.20
rotate-4ms64w1 (workloads/) 17.21
x264-4mq (workloads/) 0.41
x264-4mqw1 (workloads/) 0.41

Table: Multibench


3.1.2.2. Ethernet

Ethernet performance benchmarks were measured using Netperf 2.7.1 https://hewlettpackard.github.io/netperf/doc/netperf.html Test procedures were modeled after those defined in RFC-2544: https://tools.ietf.org/html/rfc2544, where the DUT is the TI device and the “tester” used was a Linux PC. To produce consistent results, it is recommended to carry out performance tests in a private network and to avoid running NFS on the same interface used in the test. In these results, CPU utilization was captured as the total percentage used across all cores on the device, while running the performance test over one external interface.

UDP Throughput (0% loss) was measured by the procedure defined in RFC-2544 section 26.1: Throughput. In this scenario, netperf options burst_size (-b) and wait_time (-w) are used to limit bandwidth during different trials of the test, with the goal of finding the highest rate at which no loss is seen. For example, to limit bandwidth to 500Mbits/sec with 1472B datagram:

burst_size = <bandwidth (bits/sec)> / 8 (bits -> bytes) / <UDP datagram size> / 100 (seconds -> 10 ms)
burst_size = 500000000 / 8 / 1472 / 100 = 425

wait_time = 10 milliseconds (minimum supported by Linux PC used for testing)

UDP Throughput (possible loss) was measured by capturing throughput and packet loss statistics when running the netperf test with no bandwidth limit (remove -b/-w options).

In order to start a netperf client on one device, the other device must have netserver running. To start netserver:

netserver [-p <port_number>] [-4 (IPv4 addressing)] [-6 (IPv6 addressing)]

Running the following shell script from the DUT will trigger netperf clients to measure bidirectional TCP performance for 60 seconds and report CPU utilization. Parameter -k is used in client commands to summarize selected statistics on their own line and -j is used to gain additional timing measurements during the test.

#!/bin/bash
for i in 1
do
   netperf -H <tester ip> -j -c -l 60 -t TCP_STREAM --
      -k DIRECTION,THROUGHPUT,MEAN_LATENCY,LOCAL_CPU_UTIL,REMOTE_CPU_UTIL,LOCAL_BYTES_SENT,REMOTE_BYTES_RECVD,LOCAL_SEND_SIZE &

   netperf -H <tester ip> -j -c -l 60 -t TCP_MAERTS --
      -k DIRECTION,THROUGHPUT,MEAN_LATENCY,LOCAL_CPU_UTIL,REMOTE_CPU_UTIL,LOCAL_BYTES_SENT,REMOTE_BYTES_RECVD,LOCAL_SEND_SIZE &
done

Running the following commands will trigger netperf clients to measure UDP burst performance for 60 seconds at various burst/datagram sizes and report CPU utilization.

  • For UDP egress tests, run netperf client from DUT and start netserver on tester.
netperf -H <tester ip> -j -c -l 60 -t UDP_STREAM -b <burst_size> -w <wait_time> -- -m <UDP datagram size>
   -k DIRECTION,THROUGHPUT,MEAN_LATENCY,LOCAL_CPU_UTIL,REMOTE_CPU_UTIL,LOCAL_BYTES_SENT,REMOTE_BYTES_RECVD,LOCAL_SEND_SIZE
  • For UDP ingress tests, run netperf client from tester and start netserver on DUT.
netperf -H <DUT ip> -j -C -l 60 -t UDP_STREAM -b <burst_size> -w <wait_time> -- -m <UDP datagram size>
   -k DIRECTION,THROUGHPUT,MEAN_LATENCY,LOCAL_CPU_UTIL,REMOTE_CPU_UTIL,LOCAL_BYTES_SENT,REMOTE_BYTES_RECVD,LOCAL_SEND_SIZE

3.1.2.2.1. CPSW Ethernet Driver

TCP Bidirectional Throughput

TCP Window Size am64xx-evm: Throughput (Mbits/sec) am64xx-evm: CPU Load %
Default 1070.56 97.56

Table: CPSW TCP Bidirectional Throughput

UDP Throughput (0% loss)

Frame Size(bytes) am64xx-evm: UDP Datagram Size(bytes) am64xx-evm: Throughput (Mbits/sec) am64xx-evm: CPU Load %
64 18.00 9.82 89.72
128 82.00 37.75 86.08
256 210.00 117.21 88.69
512 466.00 87.53 34.47
1024 978.00 517.58 91.11
1280 1234.00 614.84 90.96
1518 1472.00 561.33 81.58

Table: CPSW UDP Egress Throughput (0% loss)

Frame Size(bytes) am64xx-evm: UDP Datagram Size(bytes) am64xx-evm: Throughput (Mbits/sec) am64xx-evm: CPU Load %
64 18.00 1.15 11.26
128 82.00 6.56 17.32
256 210.00 16.46 8.26
512 466.00 37.65 8.46
1024 978.00 81.37 10.00
1280 1234.00 103.65 17.13
1518 1472.00 123.64 17.91

Table: CPSW UDP Ingress Throughput (0% loss)

UDP Throughput (possible loss)

Frame Size(bytes) am64xx-evm: UDP Datagram Size(bytes) am64xx-evm: Throughput (Mbits/sec) am64xx-evm: CPU Load % am64xx-evm: Packet Loss %
64 18.00 9.82 89.72 0.00
128 82.00 37.75 86.08 0.00
256 210.00 117.21 88.69 0.00
512 466.00 87.53 34.47 0.00
1024 978.00 517.58 91.11 0.00
1280 1234.00 614.84 90.96 0.00
1518 1472.00 561.33 81.58 0.00

Table: CPSW UDP Egress Throughput (possible loss)

Frame Size(bytes) am64xx-evm: UDP Datagram Size(bytes) am64xx-evm: Throughput (Mbits/sec) am64xx-evm: CPU Load % am64xx-evm: Packet Loss %
64 18.00 18.41 76.61 55.38
128 82.00 83.98 78.33 54.91
256 210.00 210.20 80.36 52.82
512 466.00 460.51 81.91 47.42
1024 978.00 922.94 84.20 1.21
1280 1234.00 943.04 84.71 0.64
1518 1472.00 942.31 83.96 1.54

Table: CPSW UDP Ingress Throughput (possible loss)


3.1.2.3. PCIe Driver

3.1.2.3.1. PCIe-NVMe-SSD

Buffer size (bytes) am64xx-evm: Write EXT4 Throughput (Mbytes/sec) am64xx-evm: Write EXT4 CPU Load (%) am64xx-evm: Read EXT4 Throughput (Mbytes/sec) am64xx-evm: Read EXT4 CPU Load (%)
1m 352.00 91.40 390.00 86.89
4m 353.00 95.64 397.00 89.48
4k 31.80 98.21 39.00 96.47
256k 276.00 95.33 339.00 84.80
  • Filesize used is: 10G
  • FIO command options: –ioengine=libaio –iodepth=4 –numjobs=1 –direct=1 –runtime=60 –time_based
  • Platform: Speed 8GT/s, Width x1
  • SSD being used: Lite-On Technology Corporation M8Pe Series NVMe SSD [14a4:22f1] (rev 01)

3.1.2.4. OSPI Flash Driver

File size (Mbytes) am64xx-evm: Raw Read Throughput (Mbytes/sec)
50 33.11

3.1.2.5. USB Driver

3.1.2.5.1. USB Host Controller

Warning

IMPORTANT: For Mass-storage applications, the performance numbers can be severely affected if the media is mounted in sync mode. Hot plug scripts in the filesystem mount removable media in sync mode to ensure data integrity. For performance sensitive applications, umount the auto-mounted filesystem and re-mount in async mode.

Setup : SAMSUNG 850 PRO 2.5” 128GB SATA III Internal Solid State Drive (SSD) used with Inateck ASM1153E enclosure is connected to usb port under test. File read/write performance data is captured.

Buffer size (bytes) am64xx-evm: Write EXT4 Throughput (Mbytes/sec) am64xx-evm: Write EXT4 CPU Load (%) am64xx-evm: Read EXT4 Throughput (Mbytes/sec) am64xx-evm: Read EXT4 CPU Load (%)
1m 37.90 7.83 38.60 7.16
4m 38.10 6.44 38.90 6.43
4k 17.60 34.65 17.70 28.36
256k 36.40 9.30 37.10 8.96