Optimize HTTP.header_name

[jgaskins]

Using Indexable#bsearch operates in logarithmic time, but Hash#[]? operates in constant time.

Benchmark code

require "set"

COMMON_HEADERS = %w(
  Accept-Encoding
  Accept-Language
  Accept-encoding
  Accept-language
  Allow
  Cache-Control
  Cache-control
  Connection
  Content-Disposition
  Content-Encoding
  Content-Language
  Content-Length
  Content-Type
  Content-disposition
  Content-encoding
  Content-language
  Content-length
  Content-type
  ETag
  Etag
  Expires
  Host
  Last-Modified
  Last-modified
  Location
  Referer
  User-Agent
  User-agent
  accept-encoding
  accept-language
  allow
  cache-control
  connection
  content-disposition
  content-encoding
  content-language
  content-length
  content-type
  etag
  expires
  host
  last-modified
  location
  referer
  user-agent
)
  .map { |header| {header.to_slice, header} }
  .to_h

# :nodoc:
def header_name(slice : Bytes) : String
  # Check if the header name is a common one.
  # If so we avoid having to allocate a string for it.
  if slice.size < 20 && (name = COMMON_HEADERS[slice]?)
    name
  else
    String.new(slice)
  end
end

require "benchmark"
require "http"

value = nil
puts "hit"
bench "allow"
puts
puts "miss"
bench "florp"
puts
puts "miss - long name"
bench "Strict-Transport-Security"

# side effect so LLVM won't optimize out the benchmark blocks
pp value unless value

def bench(header_name : String)
  Benchmark.ips do |x|
    x.report "HTTP.header_name" { value = HTTP.header_name(header_name.to_slice) }
    x.report "optimized" { value = header_name(header_name.to_slice) }
  end
end

➜  Code crystal run --release bsearch_vs_set.cr
hit
HTTP.header_name  43.35M ( 23.07ns) (± 0.37%)  0.0B/op   2.95× slower
       optimized 127.99M (  7.81ns) (± 0.63%)  0.0B/op        fastest

miss
HTTP.header_name  36.37M ( 27.50ns) (± 0.54%)  32.0B/op   1.79× slower
       optimized  65.16M ( 15.35ns) (± 2.01%)  32.0B/op        fastest

miss - long name
HTTP.header_name  74.37M ( 13.45ns) (± 0.75%)  48.0B/op        fastest
       optimized  73.93M ( 13.53ns) (± 1.10%)  48.0B/op   1.01× slower

[HertzDevil]

Just wondering, could StringPool be even faster? Something like %w(...).each_with_object(StringPool.new) { |pool, str| pool.get(str) }

[jgaskins]

Maybe, I can check it out.

[jgaskins]

Hmm, is there a way to avoid adding the string to the StringPool? We don't want to store every single header ever sent/received. That's a vector for denial of service through resource depletion.

[straight-shoota]

Not yet, but there's a feature request: #11217

[straight-shoota]

@jgaskins StringPool#get? is now available (#14508) if you want to try this.