[DNM] Experiment with always creating full FC snapshots

[maggie-lou]

                                         │ base_10_30.txt │            new_10_30.txt            │
                                         │     sec/op     │   sec/op     vs base                │
_PausePerformance/memory_size_500_mb-16       3.805 ±  2%   3.306 ± 32%        ~ (p=0.089 n=10)
_PausePerformance/memory_size_1000_mb-16      8.931 ± 15%   6.319 ± 17%  -29.25% (p=0.000 n=10)
_PausePerformance/memory_size_2000_mb-16      13.98 ±  7%   13.93 ±  1%        ~ (p=0.853 n=10)
_PausePerformance/memory_size_4000_mb-16      23.38 ±  2%   23.74 ±  5%        ~ (p=0.247 n=10)
_PausePerformance/memory_size_8000_mb-16      22.99 ±  3%   42.59 ±  1%  +85.29% (p=0.000 n=10)
geomean                                       12.06         12.41         +2.89%

[maggie-lou]

There's a huge jump in run time for memory_size_8000_mb-16. I ran it again and saw the exact same thing, so it must hit a point where exporting a very large full memory snapshot sees a large degradation in performance. Just creating a full snapshot file on an 8GB machine (firecracker code) consistently takes 35s. Creating a diff snapshot for the same size machine takes ~10s and merging takes 2-3s.

[vanja-p]

Looking at the baseline performance, it's interesting that increasing memory from 1 to 2 and 2 to 4 nearly doubles the snapshot time, but increasing from 4 to 8 actually runs faster with 8. Maybe something is up with the benchmark setup?