review of Sting

Traditional ICMP-based tools cannot measure one-way packet loss rate. Some
measurement infrastructures may require cooperation from remote endpoints.
This paper proposes the Sting which is a TCP-based network measurement
tool for one-way packet loss rates to and from unmodified hosts.

This paper discusses a loss deduction algorithm which has two phases: data
seeding and hole filling. In the data seeding phase, the algorithm sends n
data packets and counts ACKs received. In the hole filling phase, it
reliably retransmits lost packet and counts the number of lost data. It
repeats the retransmission until all packets are delivered. The forward
and reverse loss rates are calculated based on the counting information.
The algorithm is extended in three ways. First, by skipping the first
sequence number and sending data packets out-of-order, the algorithm
guarantees that one ACK is sent for each data packet received. Second,
managing limited receiver buffers by overlapping sequence numbers so that
large packets can be sent without overflow receiver buffer. Third, it
formats data packets to delay connection termination. The evaluation
results show that packet loss is highly asymmetric.

However, Sting shows that the TCP protocol features can be leveraged for
the purpose of packet loss rate measurement, maybe other purpose as well.
Sting uses a misbehaving sender to collect packet loss information. The
misbehaving client may affect the performance of the network during the
measurement. More experiments about the measurement traffic should be
executed.

The packet filters used in the Sting prototype are fragile. It may cause
connection termination in the Internet because extra copies of packets are
relayed to the TCP stack of host OS. An OS-specific firewall interface may
be a good alternative.
Received on Sun Oct 30 2005 - 20:50:53 EST