LMBench on ARM Microprocessors

What is LMBench

LMBench is an opensource benchmark suite available under the GNU General Public License. You can download LMBench here: http://sourceforge.net/projects/lmbench/

Test set up

• AM37x rev C
• OMAP EVM Main Board Rev G
• Cortex-A8 clock speed of 600MHz
• L3 clock rate of 200MHz.
• Linux version - 2.6.32 (PSP03.00.01.06)
• Compiler version - (gcc version 4.3.3 (Sourcery G++ Lite 2009q1-203)

Memory Latency

The following results come from the LMBench test lat_mem_read with stride=128. This test can be helpful to understand the latency of data reads from L1, L2 and memory. Smaller block reads can fit entirely into L1, so they should have less latency than larger block reads.

Memory Latency Results

Block tranfer size	ns	clock  cycles
512 Bytes	5.059	3.0354
2K Bytes	5.059	3.0354
4K Bytes	5.059	3.0354
16K Bytes	5.034	3.0204
32K Bytes	5.028	3.0168
48K Bytes	12.842	7.7052
64K Bytes	15.239	9.1434
128K Bytes	19.436	11.6616
192K Bytes	20.362	12.2172
256K Bytes	37.598	22.5588
384K Bytes	113.683	68.2098
512K Bytes	140.427	84.2562
768K Bytes	155.576	93.3456
1M Bytes	161.576	96.9456

Read Memory Bandwidth

The following results come from the LMBench test bw_mem. This test is using the exact same setup as the previous test however, it is run with two different Cortex-A8 clock speeds - 600MHz and 1 GHz. All the reads and writes of this test are performed by the Cortex-A8.

Memory Latency Results

Operation	Clock speed - 600MHz	Clock speed -1GHz
(rd) read 1MByte - 32 bit stride - 4 bytes at a time	386.43 MB/s	412.71 MB/s
(rdwr) read/write 1 MByte - 32 bit stride - 4 bytes at a time	274.8 MB/s	310.61 MB/s
(cp) copy 1 MByte - 32 bit stride - 4 bytes at a time	328.03 MB/s	345.01 MB/s
(fcp) copy 1 MByte no stride - 4 bytes at a time	279.77 MB/s	329.95 MB/s
bcopy copy 1 MByte - 1 byte at a time	266.42 MB/s	321.96 MB/s