RPMsg Benchmark Figures

Intro

The present page aims to measure the performance of RPMsg for different number of channels and packet and transfer sizes. This is indicative only and it can vary a lot from implementation to implementation, so use it only as a rough guide.

On the Setup section, there's examples on how to set up the benchmark and its endpoints, for different number of channels, transfer and packet sizes. Note that this can also vary in terms of device numbering, firmware filenames etc.

On the Benchmarks section, the results table is present for a number of different configurations.

Source code

Source code for rpmsg-examples located at https://github.com/analogdevicesinc/rpmsg-examples, examples are compiled with CCES studio on Windows and copied over to the ADSP-SC598 board.

Setup

Setting up the endpoints

Load firmware to SHARC(s)

Generated file names of ldr files are:

rpmsg_echo_example_Core1.ldr
rpmsg_echo_example_Core2.ldr

Load Core 1 firmware for benchmarking

cd /lib/firmware
echo stop > /sys/class/remoteproc/remoteproc0/state
echo rpmsg_echo_example_Core1.ldr > /sys/class/remoteproc/remoteproc0/firmware
echo start > /sys/class/remoteproc/remoteproc0/state

Bind channels

After firmware loaded and starter, there will be created endpoints under the:

/sys/bus/rpmsg/devices

Bind 1 channel

rpmsg-bind-chardev -p virtio0.sharc-echo.-1. -n 1 -e 288 -s 100

Bind 8 channels

rpmsg-bind-chardev -p virtio0.sharc-echo.-1. -n 8 -e 288 -s 100

Bind 32 channels

rpmsg-bind-chardev -p virtio0.sharc-echo.-1. -n 32 -e 288 -s 100

Test channel(s)

Run 1 channel

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 1000000

Run 8 channels

rpmsg-xmit-p /dev/rpmsg -n 8 -e 0 -t 1000000

Run 32 channels

rpmsg-xmit-p /dev/rpmsg -n 32 -e 0 -t 1000000

Test different transfer sizes

Run 10k transfers

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 10000

Run 100k transfers

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 100000

Run 1M transfers

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 1000000

Test different packet sizes

Set packet size 1

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 1000000 -s 1

Set packet size 32

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 1000000 -s 32

Set packet size 496

rpmsg-xmit-p /dev/rpmsg -n 1 -e 0 -t 1000000 -s 496

Benchmarks

1 Channel

PACKET SIZE	TRANSFER SIZE	Single CH speed
1	100	14593
	10000	14429
	100000	14554
	1000000	14521
	10000000	14627
2	100	29262
	10000	29015
	100000	28722
	1000000	28828
	10000000	29160
16	100	244307
	10000	233222
	100000	227375
	1000000	228350
	10000000	228385
64	100	839190
	10000	855061
	100000	833525
	1000000	824475
	10000000	821174
256	100	2611392
	10000	2498536
	100000	2314672
	1000000	2290975
	10000000	2278242
496	100	4598496
	10000	4904841
	100000	5318444
	1000000	5358093
	10000000	5402480

8 Channels

PACKET SIZE	TRANSFER SIZE	Single CH speed	Average CH Speed
1	100	15383	1923
	10000	15379	1922
	100000	15325	1916
2	100	18300	2288
	10000	28289	3536
	100000	29352	3669
	1000000	29794	3724
16	100	251359	31420
	10000	232645	29081
	100000	237663	29708
	1000000	245774	30722
	10000000	236035	29504
64	100	1514642	189330
	10000	1061882	132735
	100000	1010483	126310
	1000000	957566	119696
	10000000	997557	124695
256	100	5835056	729382
	10000	4184189	523024
	100000	3407874	425984
	1000000	3532256	441532
	10000000	2505826	313228
496	100	12892509	1611564
	10000	8742713	1092839
	100000	7078932	884867
	1000000	7143015	892877
	10000000	7184777	898097

32 Channels

PACKET SIZE	TRANSFER SIZE	Single CH speed	Average CH Speed
1	100	12464	390
	10000	12790	400
	100000	12775	399
2	100	26479	827
	10000	25364	793
	100000	25339	792
16	100	351299	10978
	10000	205900	6434
	100000	197167	6161
64	100	13584233	424507
	10000	877851	27433
	100000	753868	23558
	1000000	742513	23204
256	100	23018872	719340
	10000	8198836	256214
	100000	2925997	91437
	1000000	2768895	86528
	10000000	25052788426826	87138
496	100	44478888	1389965
	10000	33994508	1062328
	100000	6907303	215853
	1000000	6170902	192841
	10000000	6076338	189886

Notes

1. For channels=1 & packet size=1, it's easy to observer that the maximum number of packets per second is 12-15k, and the bandwidth linearly follows the increase to the maximum 496 bytes. All other (multi-channel) benchmark numbers seem to follow the same pattern, i.e. the total bandwidth stays similar, and the bandwidth per channel drops as channels are added.

2. No changes were made to tune RPMsg-lite

3. Moving Vring to DDR reduces the performance by approximately 10-20%

Benchmark Plots

Single Channel Performance

Bandwidth per channel for different number of total channels

Total Bandwidth

Total bandwidth for different number of total channels