Skip to content
Siarhei Siamashka edited this page Mar 31, 2016 · 3 revisions

64-bit tinymembench build:

tinymembench v0.4.9 (simple benchmark for memory throughput and latency)

==========================================================================
== Memory bandwidth tests                                               ==
==                                                                      ==
== Note 1: 1MB = 1000000 bytes                                          ==
== Note 2: Results for 'copy' tests show how many bytes can be          ==
==         copied per second (adding together read and writen           ==
==         bytes would have provided twice higher numbers)              ==
== Note 3: 2-pass copy means that we are using a small temporary buffer ==
==         to first fetch data into it, and only then write it to the   ==
==         destination (source -> L1 cache, L1 cache -> destination)    ==
== Note 4: If sample standard deviation exceeds 0.1%, it is shown in    ==
==         brackets                                                     ==
==========================================================================

 C copy backwards                                     :   1677.6 MB/s (0.4%)
 C copy backwards (32 byte blocks)                    :   1678.6 MB/s (1.2%)
 C copy backwards (64 byte blocks)                    :   1663.4 MB/s (1.1%)
 C copy                                               :   1752.2 MB/s (0.5%)
 C copy prefetched (32 bytes step)                    :   1237.9 MB/s
 C copy prefetched (64 bytes step)                    :   1420.2 MB/s
 C 2-pass copy                                        :   1514.5 MB/s
 C 2-pass copy prefetched (32 bytes step)             :   1098.7 MB/s
 C 2-pass copy prefetched (64 bytes step)             :    933.2 MB/s
 C fill                                               :   3731.0 MB/s
 C fill (shuffle within 16 byte blocks)               :   3730.9 MB/s
 C fill (shuffle within 32 byte blocks)               :   3730.3 MB/s
 C fill (shuffle within 64 byte blocks)               :   3731.0 MB/s
 ---
 standard memcpy                                      :   1761.4 MB/s (0.3%)
 standard memset                                      :   3732.1 MB/s
 ---
 NEON LDP/STP copy                                    :   1742.8 MB/s
 NEON LD1/ST1 copy                                    :   1755.1 MB/s (0.2%)
 NEON STP fill                                        :   3731.1 MB/s
 NEON STNP fill                                       :   2644.9 MB/s (0.2%)
 ARM LDP/STP copy                                     :   1754.3 MB/s (0.4%)
 ARM STP fill                                         :   3732.2 MB/s
 ARM STNP fill                                        :   2638.0 MB/s

==========================================================================
== Framebuffer read tests.                                              ==
==                                                                      ==
== Many ARM devices use a part of the system memory as the framebuffer, ==
== typically mapped as uncached but with write-combining enabled.       ==
== Writes to such framebuffers are quite fast, but reads are much       ==
== slower and very sensitive to the alignment and the selection of      ==
== CPU instructions which are used for accessing memory.                ==
==                                                                      ==
== Many x86 systems allocate the framebuffer in the GPU memory,         ==
== accessible for the CPU via a relatively slow PCI-E bus. Moreover,    ==
== PCI-E is asymmetric and handles reads a lot worse than writes.       ==
==                                                                      ==
== If uncached framebuffer reads are reasonably fast (at least 100 MB/s ==
== or preferably >300 MB/s), then using the shadow framebuffer layer    ==
== is not necessary in Xorg DDX drivers, resulting in a nice overall    ==
== performance improvement. For example, the xf86-video-fbturbo DDX     ==
== uses this trick.                                                     ==
==========================================================================

 NEON LDP/STP copy (from framebuffer)                 :    212.6 MB/s
 NEON LDP/STP 2-pass copy (from framebuffer)          :    206.4 MB/s
 NEON LD1/ST1 copy (from framebuffer)                 :     58.0 MB/s
 NEON LD1/ST1 2-pass copy (from framebuffer)          :     57.4 MB/s
 ARM LDP/STP copy (from framebuffer)                  :    111.5 MB/s
 ARM LDP/STP 2-pass copy (from framebuffer)           :    109.7 MB/s

==========================================================================
== Memory latency test                                                  ==
==                                                                      ==
== Average time is measured for random memory accesses in the buffers   ==
== of different sizes. The larger is the buffer, the more significant   ==
== are relative contributions of TLB, L1/L2 cache misses and SDRAM      ==
== accesses. For extremely large buffer sizes we are expecting to see   ==
== page table walk with several requests to SDRAM for almost every      ==
== memory access (though 64MiB is not nearly large enough to experience ==
== this effect to its fullest).                                         ==
==                                                                      ==
== Note 1: All the numbers are representing extra time, which needs to  ==
==         be added to L1 cache latency. The cycle timings for L1 cache ==
==         latency can be usually found in the processor documentation. ==
== Note 2: Dual random read means that we are simultaneously performing ==
==         two independent memory accesses at a time. In the case if    ==
==         the memory subsystem can't handle multiple outstanding       ==
==         requests, dual random read has the same timings as two       ==
==         single reads performed one after another.                    ==
==========================================================================

block size : single random read / dual random read
      1024 :    0.0 ns          /     0.0 ns
      2048 :    0.0 ns          /     0.0 ns
      4096 :    0.0 ns          /     0.0 ns
      8192 :    0.0 ns          /     0.0 ns
     16384 :    0.0 ns          /     0.0 ns
     32768 :    0.0 ns          /     0.0 ns
     65536 :    4.4 ns          /     7.5 ns
    131072 :    6.8 ns          /    10.4 ns
    262144 :    8.0 ns          /    11.5 ns
    524288 :   12.0 ns          /    17.6 ns
   1048576 :   77.0 ns          /   116.2 ns
   2097152 :  107.6 ns          /   146.2 ns
   4194304 :  128.9 ns          /   164.0 ns
   8388608 :  140.2 ns          /   172.4 ns
  16777216 :  147.0 ns          /   178.0 ns
  33554432 :  151.2 ns          /   181.7 ns
  67108864 :  153.6 ns          /   184.0 ns

32-bit tinymembench build:

tinymembench v0.4.9 (simple benchmark for memory throughput and latency)
 
==========================================================================
== Memory bandwidth tests                                               ==
==                                                                      ==
== Note 1: 1MB = 1000000 bytes                                          ==
== Note 2: Results for 'copy' tests show how many bytes can be          ==
==         copied per second (adding together read and writen           ==
==         bytes would have provided twice higher numbers)              ==
== Note 3: 2-pass copy means that we are using a small temporary buffer ==
==         to first fetch data into it, and only then write it to the   ==
==         destination (source -> L1 cache, L1 cache -> destination)    ==
== Note 4: If sample standard deviation exceeds 0.1%, it is shown in    ==
==         brackets                                                     ==
==========================================================================
 
 C copy backwards                                     :   1653.1 MB/s (0.7%)
 C copy backwards (32 byte blocks)                    :   1658.7 MB/s (0.9%)
 C copy backwards (64 byte blocks)                    :   1668.0 MB/s (0.4%)
 C copy                                               :   1664.2 MB/s (0.8%)
 C copy prefetched (32 bytes step)                    :   1791.4 MB/s (0.3%)
 C copy prefetched (64 bytes step)                    :   1792.8 MB/s (1.0%)
 C 2-pass copy                                        :   1261.6 MB/s
 C 2-pass copy prefetched (32 bytes step)             :   1415.6 MB/s (0.1%)
 C 2-pass copy prefetched (64 bytes step)             :   1398.2 MB/s
 C fill                                               :   3587.8 MB/s
 C fill (shuffle within 16 byte blocks)               :   3588.6 MB/s
 C fill (shuffle within 32 byte blocks)               :   3586.9 MB/s
 C fill (shuffle within 64 byte blocks)               :   3585.3 MB/s
 ---
 standard memcpy                                      :   1635.9 MB/s (0.2%)
 standard memset                                      :   3578.2 MB/s
 ---
 NEON read                                            :   1827.5 MB/s
 NEON read prefetched (32 bytes step)                 :   3367.6 MB/s (0.2%)
 NEON read prefetched (64 bytes step)                 :   3364.0 MB/s
 NEON read 2 data streams                             :   1723.5 MB/s
 NEON read 2 data streams prefetched (32 bytes step)  :   3277.6 MB/s (0.2%)
 NEON read 2 data streams prefetched (64 bytes step)  :   3299.0 MB/s (0.5%)
 NEON copy                                            :   1669.5 MB/s (0.2%)
 NEON copy prefetched (32 bytes step)                 :   1845.7 MB/s (0.5%)
 NEON copy prefetched (64 bytes step)                 :   1850.5 MB/s
 NEON unrolled copy                                   :   1662.6 MB/s (0.2%)
 NEON unrolled copy prefetched (32 bytes step)        :   2173.7 MB/s
 NEON unrolled copy prefetched (64 bytes step)        :   2186.9 MB/s
 NEON copy backwards                                  :   1667.4 MB/s
 NEON copy backwards prefetched (32 bytes step)       :   1803.6 MB/s (0.3%)
 NEON copy backwards prefetched (64 bytes step)       :   1789.2 MB/s
 NEON 2-pass copy                                     :   1450.4 MB/s
 NEON 2-pass copy prefetched (32 bytes step)          :   1567.9 MB/s
 NEON 2-pass copy prefetched (64 bytes step)          :   1569.0 MB/s
 NEON unrolled 2-pass copy                            :   1361.5 MB/s
 NEON unrolled 2-pass copy prefetched (32 bytes step) :   1623.3 MB/s
 NEON unrolled 2-pass copy prefetched (64 bytes step) :   1745.2 MB/s
 NEON fill                                            :   3715.5 MB/s (0.7%)
 NEON fill backwards                                  :   3716.0 MB/s (0.2%)
 VFP copy                                             :   1669.4 MB/s (0.4%)
 VFP 2-pass copy                                      :   1384.7 MB/s
 ARM fill (STRD)                                      :   3685.7 MB/s
 ARM fill (STM with 8 registers)                      :   3714.4 MB/s (0.2%)
 ARM fill (STM with 4 registers)                      :   3707.9 MB/s (0.2%)
 ARM copy prefetched (incr pld)                       :   1783.3 MB/s
 ARM copy prefetched (wrap pld)                       :   1772.7 MB/s
 ARM 2-pass copy prefetched (incr pld)                :   1424.0 MB/s
 ARM 2-pass copy prefetched (wrap pld)                :   1418.1 MB/s
 
==========================================================================
== Framebuffer read tests.                                              ==
==                                                                      ==
== Many ARM devices use a part of the system memory as the framebuffer, ==
== typically mapped as uncached but with write-combining enabled.       ==
== Writes to such framebuffers are quite fast, but reads are much       ==
== slower and very sensitive to the alignment and the selection of      ==
== CPU instructions which are used for accessing memory.                ==
==                                                                      ==
== Many x86 systems allocate the framebuffer in the GPU memory,         ==
== accessible for the CPU via a relatively slow PCI-E bus. Moreover,    ==
== PCI-E is asymmetric and handles reads a lot worse than writes.       ==
==                                                                      ==
== If uncached framebuffer reads are reasonably fast (at least 100 MB/s ==
== or preferably >300 MB/s), then using the shadow framebuffer layer    ==
== is not necessary in Xorg DDX drivers, resulting in a nice overall    ==
== performance improvement. For example, the xf86-video-fbturbo DDX     ==
== uses this trick.                                                     ==
==========================================================================
 
 NEON read (from framebuffer)                         :     57.0 MB/s
 NEON copy (from framebuffer)                         :     56.3 MB/s
 NEON 2-pass copy (from framebuffer)                  :     56.1 MB/s
 NEON unrolled copy (from framebuffer)                :     56.3 MB/s
 NEON 2-pass unrolled copy (from framebuffer)         :     55.6 MB/s
 VFP copy (from framebuffer)                          :    402.8 MB/s
 VFP 2-pass copy (from framebuffer)                   :    370.5 MB/s (0.2%)
 ARM copy (from framebuffer)                          :    207.5 MB/s
 ARM 2-pass copy (from framebuffer)                   :    200.8 MB/s
 
==========================================================================
== Memory latency test                                                  ==
==                                                                      ==
== Average time is measured for random memory accesses in the buffers   ==
== of different sizes. The larger is the buffer, the more significant   ==
== are relative contributions of TLB, L1/L2 cache misses and SDRAM      ==
== accesses. For extremely large buffer sizes we are expecting to see   ==
== page table walk with several requests to SDRAM for almost every      ==
== memory access (though 64MiB is not nearly large enough to experience ==
== this effect to its fullest).                                         ==
==                                                                      ==
== Note 1: All the numbers are representing extra time, which needs to  ==
==         be added to L1 cache latency. The cycle timings for L1 cache ==
==         latency can be usually found in the processor documentation. ==
== Note 2: Dual random read means that we are simultaneously performing ==
==         two independent memory accesses at a time. In the case if    ==
==         the memory subsystem can't handle multiple outstanding       ==
==         requests, dual random read has the same timings as two       ==
==         single reads performed one after another.                    ==
==========================================================================
 
block size : single random read / dual random read
      1024 :    0.0 ns          /     0.0 ns
      2048 :    0.0 ns          /     0.0 ns
      4096 :    0.0 ns          /     0.0 ns
      8192 :    0.0 ns          /     0.0 ns
     16384 :    0.0 ns          /     0.0 ns
     32768 :    0.0 ns          /     0.0 ns
     65536 :    4.4 ns          /     7.4 ns
    131072 :    6.8 ns          /    10.3 ns
    262144 :    8.1 ns          /    11.6 ns
    524288 :   18.1 ns          /    29.8 ns
   1048576 :   76.2 ns          /   115.3 ns
   2097152 :  109.2 ns          /   146.5 ns
   4194304 :  131.5 ns          /   163.7 ns
   8388608 :  143.6 ns          /   172.0 ns
  16777216 :  151.9 ns          /   177.1 ns
  33554432 :  156.9 ns          /   180.7 ns
  67108864 :  159.8 ns          /   184.0 ns 

Kernel 4.9.140-tegra #1 SMP PREEMPT Wed Mar 13 00:32:22 PDT 2019 aarch64 GNU/Linux Under xorg, no compositor active, no browser or other cpu hogs.

tinymembench v0.4.9 (simple benchmark for memory thr

==========================================================================
== Memory bandwidth tests                                               ==
==                                                                      ==
== Note 1: 1MB = 1000000 bytes                                          ==
== Note 2: Results for 'copy' tests show how many bytes can be          ==
==         copied per second (adding together read and writen           ==
==         bytes would have provided twice higher numbers)              ==
== Note 3: 2-pass copy means that we are using a small temporary buffer ==
==         to first fetch data into it, and only then write it to the   ==
==         destination (source -> L1 cache, L1 cache -> destination)    ==
== Note 4: If sample standard deviation exceeds 0.1%, it is shown in    ==
==         brackets                                                     ==
==========================================================================

 C copy backwards                                     :   2949.7 MB/s (3.8%)
 C copy backwards (32 byte blocks)                    :   3011.8 MB/s
 C copy backwards (64 byte blocks)                    :   3029.2 MB/s
 C copy                                               :   3642.2 MB/s (4.1%)
 C copy prefetched (32 bytes step)                    :   3824.4 MB/s (0.3%)
 C copy prefetched (64 bytes step)                    :   3825.3 MB/s (0.4%)
 C 2-pass copy                                        :   2726.2 MB/s
 C 2-pass copy prefetched (32 bytes step)             :   2902.6 MB/s (2.5%)
 C 2-pass copy prefetched (64 bytes step)             :   2928.3 MB/s (0.3%)
 C fill                                               :   8541.0 MB/s (0.2%)
 C fill (shuffle within 16 byte blocks)               :   8518.5 MB/s (2.1%)
 C fill (shuffle within 32 byte blocks)               :   8537.1 MB/s (0.1%)
 C fill (shuffle within 64 byte blocks)               :   8528.7 MB/s (0.2%)
 ---
 standard memcpy                                      :   3558.8 MB/s
 standard memset                                      :   8520.2 MB/s
 ---
 NEON LDP/STP copy                                    :   3633.9 MB/s (4.2%)
 NEON LDP/STP copy pldl2strm (32 bytes step)          :   1451.0 MB/s (0.3%)
 NEON LDP/STP copy pldl2strm (64 bytes step)          :   1450.9 MB/s (0.5%)
 NEON LDP/STP copy pldl1keep (32 bytes step)          :   3882.5 MB/s (3.9%)
 NEON LDP/STP copy pldl1keep (64 bytes step)          :   3884.0 MB/s (0.4%)
 NEON LD1/ST1 copy                                    :   3630.8 MB/s (0.3%)
 NEON STP fill                                        :   8537.8 MB/s
 NEON STNP fill                                       :   8544.9 MB/s (1.7%)
 ARM LDP/STP copy                                     :   3635.8 MB/s (0.3%)
 ARM STP fill                                         :   8544.8 MB/s (0.1%)
 ARM STNP fill                                        :   8549.2 MB/s (1.0%)
==========================================================================
== Framebuffer read tests.                                              ==
==                                                                      ==
== Many ARM devices use a part of the system memory as the framebuffer, ==
== typically mapped as uncached but with write-combining enabled.       ==
== Writes to such framebuffers are quite fast, but reads are much       ==
== slower and very sensitive to the alignment and the selection of      ==
== CPU instructions which are used for accessing memory.                ==
==                                                                      ==
== Many x86 systems allocate the framebuffer in the GPU memory,         ==
== accessible for the CPU via a relatively slow PCI-E bus. Moreover,    ==
== PCI-E is asymmetric and handles reads a lot worse than writes.       ==
==                                                                      ==
== If uncached framebuffer reads are reasonably fast (at least 100 MB/s ==
== or preferably >300 MB/s), then using the shadow framebuffer layer    ==
== is not necessary in Xorg DDX drivers, resulting in a nice overall    ==
== performance improvement. For example, the xf86-video-fbturbo DDX     ==
== uses this trick.                                                     ==
==========================================================================

 NEON LDP/STP copy (from framebuffer)                 :    766.0 MB/s
 NEON LDP/STP 2-pass copy (from framebuffer)          :    688.8 MB/s
 NEON LD1/ST1 copy (from framebuffer)                 :    770.6 MB/s (0.1%)
 NEON LD1/ST1 2-pass copy (from framebuffer)          :    681.3 MB/s (0.3%)
 ARM LDP/STP copy (from framebuffer)                  :    766.1 MB/s
 ARM LDP/STP 2-pass copy (from framebuffer)           :    689.1 MB/s


==========================================================================
== Memory latency test                                                  ==
==                                                                      ==
== Average time is measured for random memory accesses in the buffers   ==
== of different sizes. The larger is the buffer, the more significant   ==
== are relative contributions of TLB, L1/L2 cache misses and SDRAM      ==
== accesses. For extremely large buffer sizes we are expecting to see   ==
== page table walk with several requests to SDRAM for almost every      ==
== memory access (though 64MiB is not nearly large enough to experience ==
== this effect to its fullest).                                         ==
==                                                                      ==
== Note 1: All the numbers are representing extra time, which needs to  ==
==         be added to L1 cache latency. The cycle timings for L1 cache ==
==         latency can be usually found in the processor documentation. ==
== Note 2: Dual random read means that we are simultaneously performing ==
==         two independent memory accesses at a time. In the case if    ==
==         the memory subsystem can't handle multiple outstanding       ==
==         requests, dual random read has the same timings as two       ==
==         single reads performed one after another.                    ==
==========================================================================

block size : single random read / dual random read, [MADV_NOHUGEPAGE]
      1024 :    0.0 ns          /     0.1 ns 
      2048 :    0.0 ns          /     0.1 ns 
      4096 :    0.0 ns          /     0.1 ns 
      8192 :    0.0 ns          /     0.1 ns 
     16384 :    0.1 ns          /     0.1 ns 
     32768 :    1.7 ns          /     2.9 ns 
     65536 :    6.4 ns          /     9.5 ns 
    131072 :    9.6 ns          /    12.3 ns 
    262144 :   13.7 ns          /    17.0 ns 
    524288 :   15.8 ns          /    19.7 ns 
   1048576 :   17.3 ns          /    22.1 ns 
   2097152 :   42.1 ns          /    64.2 ns 
   4194304 :   98.5 ns          /   138.1 ns 
   8388608 :  143.9 ns          /   186.3 ns 
  16777216 :  167.2 ns          /   211.2 ns 
  33554432 :  180.1 ns          /   227.1 ns 
  67108864 :  200.0 ns          /   260.2 ns 
block size : single random read / dual random read, [MADV_HUGEPAGE]
      1024 :    0.0 ns          /     0.0 ns 
      2048 :    0.0 ns          /     0.0 ns 
      4096 :    0.0 ns          /     0.0 ns 
      8192 :    0.0 ns          /     0.0 ns 
     16384 :    0.0 ns          /     0.0 ns 
     32768 :    0.0 ns          /     0.0 ns 
     65536 :    6.4 ns          /     9.4 ns 
    131072 :    9.5 ns          /    12.2 ns 
    262144 :   11.2 ns          /    13.1 ns 
    524288 :   12.1 ns          /    13.5 ns 
   1048576 :   12.8 ns          /    13.6 ns 
   2097152 :   27.0 ns          /    33.0 ns 
   4194304 :   90.6 ns          /   127.8 ns 
   8388608 :  123.9 ns          /   153.8 ns 
  16777216 :  139.5 ns          /   161.2 ns 
  33554432 :  147.2 ns          /   163.6 ns 
  67108864 :  154.0 ns          /   167.6 ns 
Clone this wiki locally