Linux Performance Monitoring using PMU and TI EMIF data - ARM Cycles, Cache misses, Real DDR bandwidth, and more ...

--> insmod peemuperf.ko evdelay=500 evlist=1,68,3,4 evdebug=1 emifcnt=1 emiflist=2,3

--> rmmod peemuperf.ko

Parameters:

evdelay = Delay between successive reading of samples from event counters (milliSec)

evlist = Array of decimal values of event IDs to be monitored (refer ARM TRM). If not specified, first four of below are used:

1 ==> Instruction fetch that causes a refill at the lowest level of instruction or unified cache

68 ==> Any cacheable miss in the L2 cache

3 ==> Data read or write operation that causes a refill at the lowest level of data or unified cache

4 ==> Data read or write operation that causes a cache access at the lowest level of data or unified cache

67 ==> All cache accesses

0x56 ==> Increment for every cycle that no instructions are available for issue

(for other valid values, refer to Cortex-A TRM)

evdebug = 0 (default - no messages from kernel module) / 1 (event counters are printed out - warning: will flood the console)

emifcnt = 0 (default - no usage of EMIF DDR monitor) / 1 (use EMIF DDR monitor). For more details, refer to EMIF documentation for example at www.ti.com/lit/ug/sprugv8c/sprugv8c.pdf

emiflist = Array of decimal values of EMIF event IDs to be monitored. If not specified, ID 2 (READs), ID 3(WRITES) are used.

0x4 ==> Total DDR clock cycles Command FIFO is full

0x5 ==> Total DDR clock cycles Write Data FIFO is full

0xA ==> Total DDR clock cycles for which DDR i/f was transferring data

(for other valid values, refer EMIF documentation)

An example console output when evdebug = 1, is below (CPU MHz = 720 MHz, sampling every 1000 millisec): (EMIF outputs are not shown - truncated)

[ 100.066070] 49031 2388256 173679 20677707 0 11362862
[ 101.096008] 50458 2322475 169335 20240942 0 11362603
[ 102.126007] 75056 2384952 184637 20614303 0 11362506
[ 103.156036] 49902 2322703 169203 20224376 0 11362827
[ 104.186035] 48974 2320928 168278 20198781 0 11362768

The Counter values, Overflow indication, Cycle count, EMIF read and write count is provided.

NOTE: PMU Cycle count is configured to be divided by 64, compared to CPU clock

peemuperf exposes event information (same information as above) via proc entry, and can be used by userland applications

cat /proc/peemuperf

root@am335x-evm:/media/sda1# cat /proc/peemuperf
PMU.counter[0]= 54363
PMU.counter[1]= 2339356
PMU.counter[2]= 172534
PMU.counter[3]= 20313519
PMU.overflow= 0
PMU.CCNT= 11362787
EMIF.readcount= 1705266985
EMIF.writecount= 2316069476

Tested on kernel 3.2, with gcc4.5 toolchain.

Validated on AM335x (OMAP3/Beagle variants will work) - for Cortex-A8. Cortex-A9 has larger counter list, and will also work

PMU interrupts not enabled and not supported, so watchout for overflow flags manually

=== [email protected]