RPI3 swap experiments

[bob prohaska]

On Sun, Jul 01, 2018 at 01:11:35PM -0700, Rodney W. Grimes wrote:
> 
> IIRC the "delay" through a usb hub, as long as your not doing transactional
> translation (TT, ie 480 mb/s down to a 12mb/s device) is on the order of
> microseconds, this is the same weither it is a powered or un powered hub.
> 
> I doubt that any flash drive can respond faster than a spinning rust
> drive with a cache can, as the cache is DRAM with tens of nano seconds
> access time.
> 
> Flash drives DO have an access delay, the FTL still has to do the mapping
> funcions, and a page of flash has to be pulled into a read buffer.  I
> actually believe that even the fastest USB flash drives are still
> slower than any modern spinning rust at this operation.
> 
There is a sorted list of gstat output from one of my tests at
http://www.zefox.net/~fbsd/rpi3/swaptests/3gbusbflash/readdelay.sort 
The first few lines suggest that microSD responds in less than a
millisecond, direct-connected USB flash requires 1.2 milliseconds
and hub-connected USB flash takes 1.6 milliseconds. I'm not sure
how representative these numbers are, the -u option was used in sort.

There's a similar list for a mechanical drive at
http://www.zefox.net/~fbsd/rpi3/swaptests/newtests/1.3gbusbmechanical_swapinfo/readdelay.sort
where the minimum delay is 6.4 milliseconds. It's tempting to think
that at least occasionally flash storage beats mechanical storage. 

The mechanical drive I used is an old 2.5" salvaged from a Dell 
compact desktop computer, stuffed into a USB enclosure, so it
surely isn't the fastest in the world.

Thanks for reading!

bob prohaska