ben-wpan/atben/misc/atben-spi-performance.txt

Performance analysis of SPI drivers for ATBEN
=============================================

This was originally posted on April 10 2013 to the linux-zigbee and
the qi-hardware list. The URLs have been updated for stability.

Note that the code referenced in this analysis is meant to illustrate
the general concepts. Some variants have been dropped since and the
ones still in existence have been cleaned up and debugged in
preparation for upstream submission. The current versions can be
found in https://github.com/wpwrak/ben-wpan-linux


What this is all about
----------------------

The ATBEN board on the Ben NanoNote needs a bit-banging SPI driver.
There are several ways to implement this, ranging from reuse of the
generic spi-gpio driver to an optimized driver that's specific for
this platform.

I implemented several such approaches and measured their performance
in the Ben NanoNote. Below are my findings.

Comments welcome.


Cast and characters
-------------------

spi_atben_gpio: NanoNote-specific framework for setting up the
AT86RF230/1 with SPI-GPIO or one of the optimized drivers (below).
The name derives from spi_atben (see below) and should be changed
(maybe to atben_spi or atben_spi_gpio ?) since it is not an SPI
driver but merely a framework that provides configuration data and
performs miscellaneous platform setup.
https://github.com/wpwrak/ben-wpan-linux/blob/perfcomp/drivers/net/ieee802154/spi_atben_gpio.c

spi_atben: like spi_atben_gpio, but contains a highly optimized
SPI driver for the ATBEN configuration in the Ben NanoNote.
https://github.com/wpwrak/ben-wpan-linux/blob/perfcomp/drivers/net/ieee802154/spi_atben.c

spi-jz4740-gpio: SPI-GPIO driver optimized for the Jz4740. Uses the
optimized register accesses from spi_atben but pin assignment is not
restricted to ATBEN. The only limitation is that MOSI, MISO, and
SCLK must be on the same port.
https://github.com/wpwrak/ben-wpan-linux/blob/perfcomp/drivers/spi/spi-jz4740-gpio.c

spi-gpio-atben: task-specific SPI-GPIO driver using the #include
"spi-gpio.c" method. Replaces gpiolib functions with register
accesses specific to the ATBEN configuration in the Ben NanoNote.
Note that some of the code could be moved into Jz4740
architecture-specific GPIO support.
https://github.com/wpwrak/ben-wpan-linux/blob/perfcomp/drivers/spi/spi-gpio-atben.c

In the following sections, we abbreviate the stack configurations
as follows:

Abbreviation	Framework	Transport	Chip driver
---------------	---------------	---------------	-----------
spi-gpio	spi_atben_gpio	spi-gpio	at86rf230
spi-gpio-atben	spi_atben_gpio	spi-gpio-atben	at86rf230
spi-jz4740-gpio	spi_atben_gpio	spi-jz4740-gpio	at86rf230
spi_atben	spi_atben			at86rf230


Measurements
------------

Access time to AT86RF231 registers and buffer, in microseconds, on
an otherwise idle Ben NanoNote:

Driver		read from 0x51		read 120 bytes from buffer
|		|	write 0x0a to 0x15	write 1 byte to buffer (0x33)
|		|	|	read 1 byte from buffer	write 120 bytes
|		|	|	|	|	|	|
spi-gpio	 81	 85	186	1696	 97	1596
spi-gpio-atben	 63	 59	123	 498	 65	 437
spi-jz4740-gpio	 10	  8	 21	 280	 10	 231
spi_atben	 10	  7	 21	 280	 10	 230

Data rate for hypothetical buffer accesses of infinite length.
I.e., kbps = 1000*119*8/(t_write120-t_write1)

Driver		buffer read (kbps)	buffer write (kbps)
---------------	-----------------------	-------------------
spi-gpio	 630			 635			
spi-gpio-atben	2549			2559
spi-jz4740-gpio	3676			4308
spi_atben	3676			4327

At the air interface, IEEE 802.15.4 has a data rate of 250 kbps.
The AT86RF231 transceiver also supports non-standard higher data
rates up to 2 Mbps.

Driver(s)				Code size (lines)
---------------------------------------	-----------------
spi_atben_gpio				 128
spi_atben_gpio + spi-gpio-atben		 128+ 53
spi_atben_gpio + spi-jz4740-gpio	 128+416
spi_atben				 423


Computational cost
------------------

The high-level operations of sending and receiving produce the
following major low-level operations:

Operation	register	buffer		waitqueue
		read	write	read	write
---------------	-------	-------	-------	-------	---------
reception	1	-	1	-	1
transmission	9	4	-	1	1

Using the measured data from above, we get the following total
computational overhead in microseconds, without considering the
waitqueue scheduling delay:

Driver		reception		transmission
		1	120	127	1	120	125 (bytes)
---------------	-------	-------	-------	-------	-------	-------
spi-gpio	 267	1777	1866	1166	2665	2727
spi-gpio-atben	 186	 561	 583 	 868	1240	1256
spi-jz4740-gpio	  31	 290	 304 	 132	 353	 362

Note that the minimum frame length in IEEE 802.15.4 is 5 bytes.
The values for 125 (excluding CRC) and 127 (including CRC) bytes
are extrapolated.

According to [1], maximum-sized frames can be sent/received,
including CSMA/CA and acknowledgement, at a rate between one
every 4928 us and one every 7168 us.

We would therefore get the following maximum CPU load:

Driver		Reception	Transmission
---------------	---------------	------------
spi-gpio	38%		55%
spi-gpio-atben	12%		25%
spi-jz4740-gpio	 6%		 7%


Observations
------------

spi-gpio needs the smallest amount of new code but is also very
inefficient, making it questionable whether this configuration
would yield acceptable performance in regular use.

With spi-gpio-atben, only a small amount of code is added, but
buffer accesses become almost 4 times faster. Register reads and
writes are still fairly slow.

spi_atben and spi-jz4740-gpio both achieve the best performance
without significant differences between them. Both add a complete
SPI driver. Of the two, spi-jz4740-gpio is preferable, because it
uses the nearly universal spi_atben_gpio framework driver.


Conclusion
----------

I think performance trumps most other considerations in this case.
spi-gpio is clearly too inefficient. spi_atben_gpio with
spi-jz4740-gpio offers the best performance and has a low impact
on the system load (< 10%). In case this solution would be met
with strong resistance for some reason, spi-gpio-atben would offer
a compromise between performance and the amount of code.


References
----------

[1] http://www.jennic.com/files/support_files/JN-AN-1035%20Calculating%20802-15-4%20Data%20Rates-1v0.pdf