Minimizing Flow Completion Times using Adaptive

Minimizing Flow Completion Times using Adaptive Routing over Inter-Datacenter Wide Area Networks Introduction and Definitions

Microsoft Global WAN (datacenters in 38 regions)

I Throughput-oriented flows deliver large volumes of data. I Less sensitive to propagation and initial routing latency. I Inter-datacenter networks managed by one organization allow for flexible application of custom routing techniques. I Focus on single path routing and aim to minimize completion times and bandwidth usage. I Adaptive flow routing according to network and flow properties. I Fast heuristic schemes use a cost (distance) metric and select the minimum cost (shortest) path.

Path Selection Heuristics I I I I I I

Three edge cost metrics: load, load+demand, utilization load: total remaining bytes to be sent on a link (given flows on it) demand: total bytes of the new flow (its size) utilization: current fraction of capacity in use on a link Two path cost functions of MINMAX() and MINSUM() Extensively used for traffic engineering: MINMAX(utilization)

https://azure.microsoft.com/en-us/blog/ how-microsoft-builds-its-fast-and-reliable-global-network/

Cogent WAN (197 nodes and 243 edges)

Simulation Setup I Cogent WAN (uniform capacity of 1.0 for all links). I Four flow demand distributions of light-tailed, heavy-tailed, Cache-Follower and Hadoop (last two reported by Facebook). I Poisson distribution with rate λ = 1.0 for flow arrivals. I Flow demands with an average of 20 units and a maximum of 500 units (heavy-tailed had a minimum demand of 2 units). I Three scheduling policies of FCFS, SRPT and Fair Sharing using Max-Min Fairness (MMF).

Topology Info: http://www.topology-zoo.org/files/Cogentco.gml

Simulation Results

MINSUM(load) MINMAX(load) MINSUM(load+demand) MINMAX(load+demand) MINSUM(utilization) MINMAX(utilization) MinHop

MINSUM(load) MINMAX(load) MINSUM(load+demand) MINMAX(load+demand) MINSUM(utilization) MINMAX(utilization) MinHop < 10% from min

MFCT Light-tailed Heavy-tailed F S M F S M

TFCT Light-tailed Heavy-tailed F S M F S M

Total Bandwidth Used Light-tailed Heavy-tailed F S M F S M

MFCT Cache-Follower Hadoop F S M F S M

TFCT Cache-Follower Hadoop F S M F S M

Total Bandwidth Used Cache-Follower Hadoop F S M F S M

< 20% from min

< 30% from min

< 40% from min

< 50% from min

≥ 50% from min

Fig. 1. Performance of various cost metrics for path selection over Cogent WAN [6], with uniform capacity of 1 and λ = 1.0 (F , S and M represent the Figure: Performance of various costscheduling metrics for path selection oversimulation Cogent WAN (F , S and represent the FCFS, SRPT and MMF scheduling policies, respectively) FCFS, SRPT and MMF policies, respectively), was repeated manyM times and average was computed

Discussion

Future Directions

Schemes based on utilization are at least 40% above the mini- is also an effective metric for selection of multicast forwarding I Schemes based utilization are atofleast 40% above I Software Networking cantimes be used for realization of [8]. treesDefined that reduce completion via load balancing [7], mumonfor the majority scenarios. Also, the MINMAX(load) minimum for the of scenarios. MINSUM(load+demand) access to a global view of It is also interesting tohaving note that MINMAX(utilization), and majority MINMAX(load+demand) are more than 50% above network status and flow used demands. I MINMAX(load) and MINMAX(load+demand) are 50% or the minimum in mean completion times for multiple scenarwhich is frequently in traffic engineering research, is far more above minimum in mean for multiple scenarios. Distributed implementation viathe end-point switch scenarios. ios. Overall, it cantimes be seen that schemes based on I “load” from the best solution for majorityand of evaluated cooperation (flow demand encoded header). I Schemes based on “load” offermuch muchbetter bettertail tail completion times (less times (less Centralized frameworks, such in as packet SDN [9], are good candias link cost offer than 10% away for minimum majority offorcases). In case exact size is of unknown, an estimate beaccess to dates for flow realization this scheme since theycan offer thanfrom 10%minimum away from majority of cases).IAlso, used.global Further research is needed on how demand view of network status and flow flow demands. To properly MINSUM(load+demand) best mean completion I MINSUM(load+demand) offers the bestoffers meanthecompletion estimation can affect quality of selected updateaccuracy load variables associated with links, onepaths. needs knowltimes all considering all scenarios. times considering scenarios. Total Bandwidth MINSUM(load+demand) of- edge and of flow demands. In case exact flow size isforunknown, I Extending evaluating MINSUM(load+demand) I MINSUM(load+demand) offers Usage: the minimum extra fers compared the minimum extra which bandwidth usage compared multipath to an estimate be demand used. Further is needed on how routing can (what to useresearch per sub-flow). bandwidth usage to MinHop is below 20%. MinHop which is below 20% at all times. Schemes flow demand estimation accuracy can affect quality of selected based on MINMAX() consume at least 40% extra band- paths. In addition, we plan to extend and evaluate our proposed width. MINSUM(load) and use adaptive approach for multipath traffic engineering. Mohammad Noormohammadpour andMINSUM(utilization) Cauligi S. Raghavendra at of least 10% Engineering, more bandwidth at all times compared to of Southern California Ming Hsieh Department Electrical Viterbi School of Engineering, University R EFERENCES

Online

[Poster Abstract]