Only use 48 bits for encoding timestamps and 32 bits for encoding thread IDs in RawEvent in order to make it smaller.

[michaelwoerister]

This restricts us to timestamps of at most ~78 hours after process start which seems acceptable.
The PR reduces the size of RawEvent (and thus the size of our event stream files) by 25%.

[michaelwoerister]

Note to myself for tomorrow: Make event encoding endianess independent as part of this PR.

[michaelwoerister]

I added two commits:

• Store events properly in little endian format on disk.
• Optimize the FileSerializationSink by directly copying data to the writing buffer.

The second commit makes the FileSerializationSink exactly as fast as the MmapSerializationSink in the benchmarks we have. But I only tested on Linux and I also remember that the benchmarks were no good indication of performance when used in rustc. So maybe that commit should be split out into a separate PR.

[michaelwoerister]

I went ahead and moved the last commit to a separate, less urgent PR (#88). This PR is ready for review.

[wesleywiser]

Great work!

[eddyb]

Out of curiosity, was the "manual" version benchmarked on x64?
The only reason I can think of is that it might be slower to write six u32s, one at a time (as opposed to 3 u64 or one u128 and one u64), but that's possible to alleviate, but first serializing to larger integers and then writing them out.

[michaelwoerister]

IIRC the two versions were almost the same with the manual one having one instruction more (and the whole fn being something 10 instructions). I don't expect there to be much difference in practice since there is more expensive stuff going (like writing this data to disk).