tor/doc/spec/proposals/ideas/xxx-bwrate-algs.txt

# The following two algorithms 


# Algorithm 1
# TODO: Burst and Relay/Regular differentiation

BwRate = Bandwidth Rate in Bytes Per Second
GlobalWriteBucket = 0
GlobalReadBucket = 0
Epoch = Token Fill Rate in seconds: suggest 50ms=.050
SecondCounter = 0
MinWriteBytes = Minimum amount bytes per write

Every Epoch Seconds:
  UseMinWriteBytes = MinWriteBytes
  WriteCnt = 0
  ReadCnt = 0
  BytesRead = 0

  For Each Open OR Conn with pending write data:
    WriteCnt++  
  For Each Open OR Conn:
    ReadCnt++  

  BytesToRead = (BwRate*Epoch + GlobalReadBucket)/ReadCnt
  BytesToWrite = (BwRate*Epoch + GlobalWriteBucket)/WriteCnt

  if BwRate/WriteCnt < MinWriteBytes:
    # If we aren't likely to accumulate enough bytes in a second to
    # send a whole cell for our connections, send partials
    Log(NOTICE, "Too many ORCons to write full blocks. Sending short packets.")
    UseMinWriteBytes = 1
    # Other option: We could switch to plan 2 here

  # Service each writable ORConn. If there are any partial writes, 
  # return remaining bytes from this epoch to the global pool
  For Each Open OR Conn with pending write data:
    ORConn->write_bucket += BytesToWrite
    if ORConn->write_bucket > UseMinWriteBytes:
      w = write(ORConn, MIN(len(ORConn->write_data), ORConn->write_bucket))
      # possible that w < ORConn->write_data here due to TCP pushback.
      # We should restore the rest of the write_bucket to the global
      # buffer
      GlobalWriteBucket += (ORConn->write_bucket - w)
      ORConn->write_bucket = 0
 
  For Each Open OR Conn:
    r = read_nonblock(ORConn, BytesToRead)
    BytesRead += r

  SecondCounter += Epoch
  if SecondCounter < 1:
    # Save unused bytes from this epoch to be used later in the second
    GlobalReadBucket += (BwRate*Epoch - BytesRead)
  else:
    SecondCounter = 0
    GlobalReadBucket = 0
    GlobalWriteBucket = 0
    For Each ORConn:
      ORConn->write_bucket = 0



# Alternate plan for Writing fairly. Reads would still be covered
# by plan 1 as there is no additional network overhead for short reads,
# so we don't need to try to avoid them.
# 
# I think this is actually pretty similar to what we do now, but 
# with the addition that the bytes accumulate up to the second mark
# and we try to keep track of our position in the write list here
# (unless libevent is doing that for us already and I just don't see it)
#
# TODO: Burst and Relay/Regular differentiation

# XXX: The inability to send single cells will cause us to block
# on EXTEND cells for low-bandwidth node pairs..
BwRate = Bandwidth Rate in Bytes Per Second
WriteBytes = Bytes per write
Epoch = MAX(MIN(WriteBytes/BwRate, .333s), .050s)

SecondCounter = 0
GlobalWriteBucket = 0

# New connections are inserted at Head-1 (the 'tail' of this circular list)
# This is not 100% fifo for all node data, but it is the best we can do
# without insane amounts of additional queueing complexity.
WriteConnList = List of Open OR Conns with pending write data > WriteBytes
WriteConnHead = 0

Every Epoch Seconds:
  GlobalWriteBucket += BwRate*Epoch
  WriteListEnd = WriteConnHead

  do
    ORCONN = WriteConnList[WriteConnHead]
    w = write(ORConn, WriteBytes)
    GlobalWriteBucket -= w
    WriteConnHead += 1
  while GlobalWriteBucket > 0 and WriteConnHead != WriteListEnd

  SecondCounter += Epoch
  if SecondCounter >= 1:
    SecondCounter = 0
    GlobalWriteBucket = 0