Re: [isabelle] Isabelle2021-RC5: support for Apple Silicon

Makarius <makarius at sketis.net> · Sat, 13 Feb 2021 13:49:20 +0100

On 08/02/2021 23:09, Makarius wrote:
> On 08/02/2021 22:42, Makarius wrote:
> 
>> Presumably the last release candidate is
>> https://isabelle.sketis.net/website-Isabelle2021-RC5
> 
> The announcement now says:
> 
>   * Support for macOS Big Sur on Intel and Apple Silicon (ARM).
> 
> For RC5 the main change is "Improved support for Apple Silicon (ARM): external
> processes are managed by Isabelle/Scala instead of Apple's Rosetta (which does
> not support multithreaded process fork)."
> 
> Thus sledgehammer works properly, and as far as I can see everything else as well.

Here are side-by-side measurements of the last two generations of Mac Minis.

Overall setup:

* Isabelle2021-RC5 (approximately, with minor/irrelevant variations in the
changeset ids)

* AFP/9a6ed530ce81 (afp-2021 branch)

* ISABELLE_TOOL_JAVA_OPTIONS="-Djava.awt.headless=true -Xms512m -Xmx4g -Xss16m"

* ML_PLATFORM="x86_64_32-darwin"
  ML_OPTIONS="--minheap 1500 --maxheap 10g"

* isabelle build -o threads=4 -d '$AFP' -f -g slow

* Machines:

  (1) Intel: Macmini8,1, i7 CPU with 6/12 cores, 64 GB RAM, SSD storage, macOS
11.2

  (2) Apple Silicon: Macmini9,1, M1 CPU with 4/8 cores, 16 GB RAM, SSD
storage, macOS 11.2.1

The counting of CPU cores has become more difficult than before: Intel has its
traditional notion of hyperthreading, but Apple M1 has a different notion of
big vs. little cores. Poly/ML reports 6 cores on (1) and 8 cores on (2).

For the measurements, I have enforced 4 worker threads in Isabelle/ML and
restricted the heap size such that the Macmini9,1 still feels comfortable: it
has only 16 GB (at maximum), while the older Intel Mac Mini has 64 GB (at
maximum).

The results for AFP "slow" are included in the attachment. The new Apple
Silicon machine is slightly faster, even with the runtime translation of
Poly/ML x86_64 code to ARM64 via Rosetta 2.

Some notable results:

intel.log:Finished HOL (0:02:53 elapsed time, 0:07:52 cpu time, factor 2.72)
apple.log:Finished HOL (0:02:45 elapsed time, 0:07:22 cpu time, factor 2.67)

intel.log:Finished AODV (1:14:08 elapsed time, 4:06:45 cpu time, factor 3.33)
apple.log:Finished AODV (0:58:25 elapsed time, 3:16:21 cpu time, factor 3.36)

intel.log:Finished Flyspeck-Tame-Computation (2:38:30 elapsed time, 3:58:26
cpu time, factor 1.50)
apple.log:Finished Flyspeck-Tame-Computation (2:36:44 elapsed time, 3:56:57
cpu time, factor 1.51)

intel.log:Finished ConcurrentGC (0:13:35 elapsed time, 0:51:12 cpu time,
factor 3.77)
apple.log:Finished ConcurrentGC (0:12:27 elapsed time, 0:46:32 cpu time,
factor 3.74)

In conclusion, I am pleased to declare Apple Silicon M1 (and macOS 11 Big Sur)
as fully supported by the Isabelle2021 release (which is to be published
towards the end of next week).

Side-remark: This rather quick adoption of quite different new hardware has
been possible thanks to a generous donation of a Macmini9,1.

	Makarius