Raspberry Pi 4 vs Pi 3 performance test

[han59]

Maybe of interest,  a practical performance test between the Raspberry Pi 3 and Pi 4

Both  system where running  Raspbian Buster. The camera an ASI1600-MM Cool was connected using an USB 3.0 cable. The  imaging program was CCDCiel and  ASTAP was installed as the solver as described here.  PHD2 was installed for guiding but not used in this test.

The results show that the Pi 4 is in average twice as fast as the Pi 3. A lot of time is lost by downloading and saving the image. A class 10 memory card was installed.
The time lost is the time between the exposures for downloading the image via USB and saving it to the memory card.  So in the Pi 4  if you operate the ASI1600 camera in bin 2x2 mode there is 2.5 seconds lost between the exposures.  That's pretty good for deep sky imaging.

The solving time is the time required for plate solving the image.

Han

Edited July 12, 2019 by han59

[pete_l]

That's very interesting Han. Thank you for doing this work.
Looking at the time lost between exposures, do you think the speed difference between USB2 and USB3 is a factor here?

On a Pi4 with a lot of memory there is the option to create a RAMDISK and save the image there. That will remove any time spent saving the image to the SD card.

Edited July 12, 2019 by pete_l

[han59]

The speed difference of the USB3.0 is significant.  I assume you could  also gain with a better memory card but the 2.5 seconds are fine for me.

The RAMDISK could be interesting. This a technology used in the past. But  CCDCiel has already the image in memory, if it could start a new thread and save the image in parallel while starting the new exposure this saving doesn't count anymore.  Maybe it is already done.  I could redo the test with only viewing the image and see how much time that takes. If it is shorter, that means time can be saved by starting a new separate thread. If so we could ask the CCDCiel author Patrick Chevally  to implement a new thread for saving the image.

But in any case these "lost times" are  already pretty good.

Han

[han59]

I did some experimenting. CCDCiel has a temporary and capture folder.  Created a ram disk and set the path's to the ram disk. No performance improvement. Dropped the question to Patrick, here.

[stash_old]

12 hours ago, pete_l said:

option to create a RAMDISK and save the image there

RPI Forum did tests on Zram and others and found very little gain by using such options - although my testing on a Zero(w) did show some improvement.

[stash_old]

13 hours ago, han59 said:

The time lost is the time between the exposures for downloading the image via USB and saving it to the memory card

Maybe using a SSD via the USB3 may improve saving times and possible Solving times. BUT at the moment you can't(lat time I looked) boot from USB on RPI4.

Still good to get comparison.

Note transfer 108mb fits image file from RPI3 to AMD64 takes under 10 secs over 100mb network - as a straight copy and paste. I am not sure but transfering data between Indi and a client maybe faster than Indigo even though they use the same protocol but Indi I believe use a different compression method

[han59]

Transport & saving of images while the next exposure is running will solve all delays except the transport time from camera to Raspberry Pi.

Patrick has already created the code for an experimental CCDCiel version which does multitasking. So starting the next exposure and saving the previous image goes in parallel. First test indicates the lost time between exposures will be half or less.

Note the above test results where for local ram-disk option. Transport via INDI localhost where 10% or 15% slower.