process walker perfs: optimize readLimits and readStats

[alban]

• optimize readStats
• optimize readLimits
• cache limits and cmdline/name after parsing

Fixes #2461

[2opremio]

@alban I don't doubt that the code is much faster now but it's very hard to read. Please try to refactor it. I think that separating the parsing of integers from the space-counting code would already be an improvement.

Also, are you sure that parsing ints manually provides an improvement? I would be surprised if the Golang library wasn't optimized in this respect.

[alban]

The issue with strconv.Atoi or strconv.ParseInt is that they don't stop parsing when finding a non-digit such as a space and they return an error.

The previous code was using strings.Fields to allocate and build a string compatible with strconv.Atoi (containing only the digits without extra spaces). The pprof graphs showed that it was strings.Fields taking the cpu rather than strconv.Atoi but we cannot use Atoi without Fields.

fmt.Sscanf("%d "...) is able to work with extra spaces but it's also slower.

I'll try to refactor this.

[alban]

Code refactored and rebased.

Now that #2456 is merged in master, I ran performance tests again to see if this PR was still bringing better perfs. I used the test scenario described in #2461 (comment)

Master branch

Up to 95660e3 (so #2456 is included)

This PR

Conclusion

The function github.com/weaveworks/scope/probe/process.(*walker).Walk takes respectively:

• 15.32s
• 12.91s (16% faster)

Overall, the graphs' titles say respectively:

• 36.37s
• 31.32s (19% faster overall, probably coincidental)