Midterm Reports of MPTCP-Related Middlebox Behavior
Michio Honda, Keio University
Yoshifumi Nishida, Dyyno.Inc / WIDE project
Costin Raiciu, UCL
Mark Handley, UCL
MPTCP WG, Nov.10, 2010
79th IETF@Beijing
Today’s Data
Now we got data from more than 100 paths from around 20 countries
Today we show some of data in 81 paths obtained between 25 September and 22 October, 2010 36 (44.4%) home networks (Home-gateway+ISP) 12 (14.8%) university networks 11 (13.6%) cellular career (EDGE, 3G) 11 (13.6%) hotspots (e.g., hotel, public Wifi, hall) 6 (7.4%) company networks 4 (4.9%) hosting agents (e.g., Amazon EC2)
Middlebox Behavior MPTCP Should Care
TCP Options MP_CAPABLE, MP_DATA, MP_ACK, TIMESTAMP
Passed, removed, modified, zeroed (and at outgoing or incoming)?
Retransmission and Gapped Sequence When MPTCP retransmits the segment on another path,
retransmitting “TCP-level sequence number” might have different payload
Or, original sequence number might be skipped
TCP Options
Port 34343 3.7 % (3 out of 81) remove MP_CAP
All of them are cellular networks The same 3.7 % also remove MP_DATA
Port 80 12.3 % (10 out of 81) remove MP_CAP
4 cellular networks, 3 university networks, 3 hotspots Include all paths removing MP_CAP at port 34343
8.6 % (7 paths) remove MP_DATA Part of paths removing MP_CAP (3 cellular and 2 university 3
hotspots) 3.7 % (3 paths) didn’t deliver non-HTTP payload
The other part of paths removing MP_CAP (1 cellular, 1 univ.)
TCP Options (cont.)
Port 443 6.2 % (5 out of 81) remove MP_CAP
3 hotspots and 2 cellular networks All of these remove MP_CAP at port 80
The same 6.2 % also remove MP_DATA
Summary None of paths drops segments including MP_CAP or
MP_DATA 3.7 – 12.3 % of paths (depending on the port number)
remove MP_CAP or MP_DATA
Retransmission
Retransmission
Retransmissions with Different Payload
Port 34343 1.2 % (1 cellular network out of 81) discarded
retransmission of different payload Ack is advanced, but retransmitted payload has not reached the
receiver, because middlebox retransmitted
Port 80 7.4 % (6 out of 81) discarded retransmitted segment
including different payload, but Ack is advanced 3 hotspots, 2 cellular networks, and 1 home network One of them also does the same behavior at port 34343
1.2 % (1 out of 81) sent back RST against different payload retransmission 1 company network
Retransmissions with Different Payload (cont.)
Port 443 3.7 % (3 out of 81) discarded the retransmitted, but Ack is
advanced 3 hotspots Part of paths doing the same behavior at port 80
Summary 1.2 – 7.4 % of paths do not deliver retransmitted
segment, and the middlebox retransmits instead 1.2 % sends back RST against retransmission of different
payload for port 80
Retransmission
Gapped Sequence Number
Port 34343 1.2 % (1 cellular network out of 81) discarded the gapped
sequence 1.2 % (1 cellular network) replied Ack to resend the
correct sequence number Port 80
4.9 % (4 out of 81) discarded the gapped sequence 2 cellular networks, 1 company network and 1 home network
6.2 % (5 out of 81) replied Ack to resend the correct sequence number 3 hotspots and 2 cellular networks
Gapped Sequence Number (cont.)
Port 443 1.2 % (1 out of 81) discarded the gapped sequence
1 cellular network that also discards such segment at port 80 4.9 % (4 out of 81) replied Ack to resend the correct
sequence number 3 hotspots and 1 cellular networks (part of paths doing the
same behavior at port 80)
Summary 1.2 – 4.9 % of paths discard the gapped sequence 1.2 – 6.2 % of paths respond to resend the correct
sequence
Thank You for Your Contribution
We’re still collecting data, please send us log files from our experimental tool http://www.ietf.org/mail-archive/web/multipathtcp/curr
ent/msg01150.html If anybody stays in the hotel other than Shangri-la,
please run it at your room! If anybody can connect to Chinese 3G career via
iPhone (MyFi) or Android tethering, let us know!
Top Related