Improving datacenter performance and robustness with multipath TCP

By Costin Raiciu, Sébastien Barré, Christopher Pluntke, Adam Greenhalgh, Damon Wischik, Mark James Handley. Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication, 2011

838 citations recorded by the source corpus.

Subfields: Computer Networks and Communications, Information Systems

View Paper: https://www.semanticscholar.org/paper/925b53922da6d922b29a253ad5e31f50a29f2de7

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Networks and Communications

In Information Systems

Citations in the corpus

Resource Management in Clouds: Survey and Research Challenges (2015, background): “[179] and D(3), a deadline-aware control protocol [241].”
Presto: Edge-based Load Balancing for Fast Datacenter Networks (2015, methodology): “Negative numbers imply shorter FCT. 7.5x worse ( ∼ 11ms) and MPTCP experiences timeout (because of higher loss rates and the fact that small sub-ﬂow window sizes from small ﬂows can increase the chances of timeout [53]).”
pFabric: minimal near-optimal datacenter transport (2013, background): “Better load balancing of course reduces hotspots and thus helps reduce flow completion time, however the techniques and goals are orthogonal and complementary to pFabric.”
A measurement-based study of MultiPath TCP performance over wireless networks (2013, background): “[23, 32] showed that MPTCP outperforms standard TCP when path diversity is available in a data center network as well as in very simple wireless settings.”
CONGA: distributed congestion-aware load balancing for datacenters (2015, methodology): “However, it is well known [2, 41, 9, 27, 44, 10] that ECMP can balance load poorly.”
Re-architecting datacenter networks and stacks for low latency and high performance (2017, methodology): “Previous work has shown that the permutation traffic matrix has similar macroscopic behavior for 128, 1024 and 8192node Fat Trees [31] when 100 packet buffers are used in conjunction with Multipath TCP.”
The only constant is change: incorporating time-varying network reservations in data centers (2012, methodology): “These networks typically use hash-based or randomized techniques such as ECMP and Valiant Load Balancing to spread traf.c across multiple equal cost paths, and can use more involved techniques such as Hedera [8] and MPTCP [31] to ensure uniform traf.c spread despite .ow length variations.”
IncBricks: Toward In-Network Computation with an In-Network Cache (2017, background): “Second, datacenter networks offer many paths between end-points [24, 39, 44], and network component failures are commonplace, so keeping computations and state coherent across networking elements is complex.”
Finishing flows quickly with preemptive scheduling (2012, background): “Several recent works [17, 21] show the benefits of multipath TCP, ranging from improved reliability to higher network utilization.”
DeTail: reducing the flow completion time tail in datacenter networks (2012, background): “This phenomena has been observed before for large flow sizes [12, 29].”
Reproducible network experiments using container-based emulation (2012, background): “Students chose a wide range of projects, covering transport protocols, data center topologies, and queueing: MPTCP [48, 39], DCTCP [14], Incast [43], Outcast [40], RED [47], Flow Completion Time [45], Hed-era [13], HULL [38], Jellyﬁsh [41], DCell [44], CoDel [37], TCP Initial Congestion Window…”
RDMA over Commodity Ethernet at Scale (2016, background): “For TCP in best-eﬀort networks, there are MPTCP [29] and per-packet routing [10] for better network utilization.”
HULA: Scalable Load Balancing Using Programmable Data Planes (2016, background): “Modiﬁed transport layer: MPTCP [38] is a modiﬁed version of TCP that uses multiple subﬂows to split trafﬁc over different paths.”
How Hard Can It Be? Designing and Implementing a Deployable Multipath TCP (2012, background): “to find and utilize unused network capacity in redundant topologies[10, 19].”
MPTCP Is Not Pareto-Optimal: Performance Issues and a Possible Solution (2013, methodology): “We implemented OLIA in the simulator and use the same scenarios as [7].”
Experimental evaluation of multipath TCP schedulers (2014, background): “This enables the pooling of resources as the paths may go over different interfaces with distinct bottlenecks, effectively increasing the goodput for the end user [14, 3].”