Profile and optimize new VM.validate_header()

[gsalgado]

That method takes 1.4s to complete when I call it passing the DAO fork header (1920000). If it takes as long to validate any header, that will be a huge problem for the light client and fast-sync

Need to profile it and optimize. Unless, of course, it's only when validating the DAO fork header that it takes that long

[carver]

If 1.4s is the time to run only one test, it might be the cache generation for the PoW check. I would want to see the average time over a bunch of runs.

[gsalgado]

This is some profiling data I gathered from importing ropsten blocks around 107178. Header validation (mainly pow checking) consumes a significant amount of time, but doesn't seem to be a huge deal on that range of blocks. It'd be interesting to do the same on recent blocks, with higher TD. If anybody has a fully synced light DB, this is the code I used to profile just the header importing: https://gist.github.com/gsalgado/e8e6d3e28ad7fa3db84040be551b6bce

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
   398513    5.311    0.000  243.677    0.001 base_events.py:1336(_run_once)
   782090    2.972    0.000  148.824    0.000 events.py:143(_run)
    49536    0.049    0.000  119.393    0.002 lightchain.py:159(_process_announcements)
    49536    0.041    0.000  119.343    0.002 lightchain.py:230(process_announcement)
    49536    0.649    0.000  119.303    0.002 lightchain.py:246(_import_headers_from_peer)
    49536    0.327    0.000  114.036    0.002 lightchain.py:275(_validate_header)
    24768    0.268    0.000  110.345    0.004 base.py:676(validate_header)
    24768    0.194    0.000  109.906    0.004 base.py:703(validate_seal)
    24768    0.586    0.000  102.144    0.004 pow.py:55(check_pow)
    24768   98.904    0.004   98.904    0.004 {built-in method pyethash.hashimoto_light}

[carver]

I heard a rumor that geth doesn't check the PoW on every header. Would be worth checking out how they do that safely.

[gsalgado]

Indeed it doesn't; I describe their approach in #330 (comment). Filed #932 for it

[gsalgado]

So, as shown by the profiling data above, doing PoW validation makes light sync much slower, but it's mainly because of pyethash.hashimoto_light, which I imagine won't be easy to optimize, so I think our best bet is #932

[carver]

Should we:

1. move this issue to the following milestone, or
2. add kth header validation to Grace Hopper

I'm currently pro 1

[gsalgado]

I guess we could do both. 2 is already on its way (#1010)

[carver]

With #1010 merged, this issue seems done enough to close.