Funding
Shared Bottleneck Detection and Response For Concurrent Multipath Transfer (CMT)
PI: Janardhan Iyengar (Franklin & Marshall College)
Sponsor: Cisco Systems (RFP-999)
Amount: $93,636
Period: September 2007 - August 2008
Current Projects
Shared Bottleneck Detection and Response Techniques for CMT
My dissertation work at the University of Delaware extended the Stream Control Transmission Protocol (SCTP), a transport layer protocol, to support end-to-end Concurrent Multipath Transfer (CMT) between multihomed source and destination hosts. CMT is the concurrent transfer of new data from a source to a destination host via two or more end-to-end paths. My dissertation work operated under the assumption that the multiple paths used in CMT did not share a network bottleneck (i.e., points of congestion in the network). This research effort relaxes my assumption and investigates shared bottleneck detection and response mechanisms for CMT. We are currently looking into (i) CMT behavior in the presence of a shared bottleneck, (ii) mechanisms for a CMT sender to share congestion state when a shared bottleneck is detected, and (iii) mechanisms to seamlessly migrate between using shared and separate congestion state during an association, as per need.
Collaborators:
Owen Raccuglia (Connecticut College), Randall Stewart (Cisco Systems)
Employing SCTP's Services in SIP Networks
The Session Initiation Protocol (SIP) is an application-layer signaling protocol for communications with one or more participants. These sessions include VoIP calls, multimedia conferences, instant messaging, and others. SIP is a promising new protocol with new demands on its transport protocol. The Stream Control Transmission Protocol (SCTP) has been considered a good match for SIP transport, but much work needs to be done in understanding how SIP networks can benefit from SCTP. Our goal in this project is to identify meaningful ways in which a SIP network can employ SCTP's multistreaming service. As a first step, we have contributed SCTP support for OpenSER, a widely used open-source SIP proxy. We are currently in the process of measuring performance and implementing our architecture in OpenSER for employing multistreaming.
Collaborators:
Glenn Marmon (Connecticut College), Arup Acharya (IBM TJ Watson Research Center), Vijay Balasubramaniyan (Georgia Tech)
Exploiting Multiple Network Interfaces in Cluster Nodes for MPI
Many computing clusters use inexpensive Gigabit Ethernet, and the nodes are often equipped with multiple interfaces cards to improve bandwidth and enhance fault tolerance. In this work, we are investigating the use of Concurrent Multipath Transfer (CMT) with MPI programs in cluster networks. We have implemented MPI over CMT, and are evaluating our scheme against other current systems.
Collaborators:
Brad Penoff, Mike Tsai, Alan Wagner (Univerisity of British Columbia, BC, Canada)
Acknowledgment Congestion Control for TCP
While a Transmission Control Protocol (TCP) data sender controls its sending rate to avoid/reduce congestion on the forward path, any congestion caused by the resulting acks on the reverse path is ignored by both TCP endpoints. This uncontrolled ack rate can hurt other flows sharing the bottleneck link on the reverse path, and can also reduce the sending rate of this TCP connection due to fewer acks making it through to the sender. In this research effort, we propose an end-to-end acknowledgment congestion control mechanism for TCP that uses participation from both TCP endpoints to control the ack rate while maintaining the feedback loop to the sender.
Collaborators:
Andres Arcia, David Ros (ENST Bretagne, France), Sally Floyd (ICSI Center for Internet Research)
Past Projects
Potentially Failed State for CMT (CMT-PF) (2006-2008)
We proposed CMT with a Potentially-failed destination state (CMT- PF) to alleviate receive buffer blocking (see below). In this work, we found that by introducing this simple state, CMT-PF performs on par or much better than CMT during more aggressive failure detection thresholds than recommended by RFC4960. We also found that CMT-PF performs as well as CMT during non-failure scenarios, and interestingly, outperforms CMT when the paths experience asymmetric receive buffer blocking conditions. Based on this work, we recommended that CMT be replaced by CMT-PF in future CMT implementations and RFCs
Collaborators:
Preethi Natarajan, Paul Amer (University of Delaware), Randall Stewart (Cisco Systems)
Using SCTP For Web Transport (2005-2006)
Although TCP has traditionally been used, we argued that SCTP's services better match the needs of HTTP-based network applications and can enhance web transfers. We modified the Apache server and the Firefox browser to benefit from SCTP, and demonstrated our HTTP over SCTP design via simple experiments.
Collaborators:
Preethi Natarajan, Paul Amer (University of Delaware), Randall Stewart (Cisco Systems)
Performance Implications of a Bounded Receive Buffer in Concurrent Multipath Transfer (2005-2006)
We studied the performance of CMT in the presence of a bounded receive buffer (rbuf). We demonstrated that if two paths are used for CMT, the lower quality (i.e., higher loss rate) path degrades overall throughput of an rbuf-constrained CMT association by blocking the rbuf. We argued that rbuf blocking is not specific to the transport layer, but applies to multipath transfers at other layers as well. We investigated CMT performance under different retransmission policies, and rbuf sizes, and showed that when large differences exist in path delays and loss rates, using only the better path outperforms using two paths concurrently. While rbuf blocking cannot be eliminated, we found that it can be reduced by choice of retrans- mission policy---a mechanism available to only the transport layer. We recommended the loss-rate-based policies, which are the best performing ones, for CMT. This research was part of my dissertation work at the University of Delaware.
Collaborators:
Paul Amer (University of Delaware), Randall Stewart (Cisco Systems)
Concurrent Multipath Transfer Using SCTP Multihoming (2003-2006)
Transport layer multihoming binds a single transport layer association to multiple network addresses at each endpoint, thus allowing the two end hosts to communicate over multiple network paths. This project investigated end-to-end Concurrent Multipath Transfer (CMT) using transport layer multihoming for increased application throughput. CMT is the simultaneous transfer of new data from a source host to a destination host via two or more end-to-end paths. We investigated and evaluated design considerations in implementing CMT at the transport layer using the Stream Control Transmission Protocol (SCTP) as an example of a multihome-capable transport layer protocol. Specifically, we explored algorithms for CMT at the transport layer and retransmission policies for CMT. This research was part of my dissertation work at the University of Delaware.
Collaborators:
Paul Amer (University of Delaware), Randall Stewart (Cisco Systems)
Making SCTP More Robust to Changeover (2001-2003)
We uncovered and explored the problem of cwnd overgrowth with SCTP's currently specified handling of changeover. We developed an analytical model to abstractly quantify this cwnd overgrowth and to provide insight into the ambient conditions under which cwnd overgrowth can be observed. We used analytical results to suggest that the problem might not be a corner case. Of the various solutions that we proposed, we recommended the addition of the Split Fast Retransmit (SFR) algorithm to SCTP.
Collaborators:
Armando Caro, Paul Amer (University of Delaware), Randall Stewart (Cisco Systems)
Multistreaming and Preferential Treatment (2002)
We designed an SCTP mechanism to provide the application with the service of being able to mark data such that different parts of a transfer (different streams) could be requested to receive preferential treatment from the network. We implemented this design in the University of Essen's userland SCTP stack. A number of experiments showed that this design was reasonable. We also noted that the complexity was driving us towards separating the different streams into different associations.
Collaborators:
Sunil Samtani (Telcordia Technologies)