• Classification results
  • Initial results on top 100 websites
  • Results with the eval trick on top 1000 websites

Classification results

Initial results on top 100 websites

Of the 40116 scripts we analyzed (3192.7MB, details in script_features.py):

• Size: 9B~8.7MB, average 80kB, median 2.2kB.
  • Many small scripts.
• No significant correlation between anything.
• Coverage: 67% classified by count, 93% by size.
• Coverage growth potential: 7% by count, 1% by size.
• Half (1296.1MB, 40.60%) scripts by size fall into all sure categories. Count 1473 (3.67%).
  • Look for bloated sites.
  • What next if we can split it up?
    • What user impact from size? Aggregate per page?
    • Chrome execution time of script.

Results with the eval trick on top 1000 websites

We crawled the top 1000 websites with the eval trick applied to scripts. The eval trick rewrites each script into smaller blocks of eval calls, creating V8 execution contexts that maps to smaller blocks of source code. The unit here is execution context, not script anymore.

• Size: 1B~6MB, average 1.2kB, median 48B, total 26.2GB.
  • Most scripts broken down, but some persisted.
  • Rewritten contexts cover 23.3GB (89.14%) of source code.
    • Some rewritten contexts are still large, probably because of class, else branch, etc.
• Still no significant correlation between anything.
• Half of the contexts by size do not call any APIs, 84% by count.
• Coverage: 86% classified by count, 87% by size. By size, 49.12% silent + 38% classified + 12.87% no-sphere.
  • By size, 98% (3.3GB) of no-sphere contexts make less than 100 API calls.
  • By size, 80% (667MB) of no-sphere contexts are rewritten.
• One eighth of contexts by size fall into all sure categories (12.37%, 3235.2MB), which is 1/3 of all contexts belonging to some sphere.
  • Much less than previous results, but still high.