1. Big Data + SDN SDN Abstractions

2. The Story Thus Far Different types of traffic in clusters Background Traffic – Bulk transfers – Control messages Active Traffic (used by jobs) – HDFS read/writes – Partition-Aggregate traffic

3. The Study Thus Far Specific communication patterns in clusters – Patterns used by Big Data Analytics – You can optimize specifically for theses Map Reduce HDFS Map HDFS Reduce Shuffle Broadcast Incast

4. The Story Thus Far Helios, Hedera, MicroTE, c-thru improve utilization – Congestion leads to bad performance – Eliminate congestion Gather Network Demand Gather Network Demand Determine paths with minimal congestion Install New paths

5. Draw Backs Demand gather at network is ineffective – Assumes that past demand will predict future – Many small jobs in cluster so ineffective May Require expensive instrumentation to gather – Switch modifications – Or endhost modification to gather information

6. Application Aware Networking Insight – Application knows every the network need So application can in fact instruct the network – Small number of big data paradigms So only a small number of applications need to be changed Map-reduce/hadoop, sharp, dyrad – Application has a central entity controlling everything

7. Important Questions What information do you need? – Size of a flow – Source+destination of the flow – Start time of the flow – Deadline of the flow How should the application inform the network? – Reactively or proactively – Modified applications or unmodified applications

8. Challenges Getting Information Flow Size Insight – Data that is transferred is data that is stored in a file Input data – Query HDFS for file size Intermediate data/Output Data – Reactive methods: wait for map to finish writing to temp file Asking the file system for size Checking the Hadoop logs for file size Checking the Hadoop web-API – Proactive methods: predict size using prior history Jobs run the same code over and over Learn the ratio between input data and intermediate data Learn the ratio between intermediate data and output data

9. Challenges Getting Information End points Reactively – Job tracker places the task; it knows the locations Check the hadoop logs for the locations Modify the job tracker to directly inform you of location Proactively – Have the SDN controller tell the job tracker where to place the end-points Rack aware placement: reduce inter-rack transfers Congestion aware placement: reduce loss

10. Challenges getting information :Flow start time Hadoop specific details obscure the start time – Reducer transfers data from only 5 map at at time Tries to reduce unfairness – Reducers randomly pick the mappers to start from – Reducers start transfer at random times Tries to reduce incast – and synchronization between flows Logs store when transfer starts

11. FloxBox: Simple Approach Insight: many types of traffic exist in N/W – We only care about map-reduce more than other traffic Solution: prioritize map-reduce traffic – Place them highest priority queue – Other traffic can’t interfere How about control messages? – Should prioritize those too.

12. Reactive Approach: FlowComb Reactive attempt to integrate bigdata + SDN – No changes to application – Learn information by looking at logs and determine file size and end-points – Learn information by running agents on the endhost that determines start times

13. FlowComb: Architecture Agents on servers – Detect start/end of map – Detect start/end transfer Predictor – Determines size of intermediate data Queries Map Via API – Aggregates information from agents sends to scheduler

14. FlowComb: Architecture Scheduler – Examines each flow that has started – For each flow what is the ideal rate – Is the flow currently bottlenecked? Move to the next shortest path with available capacity

15. Open Questions How about non map-reduce traffic? – Only focus on the active transfers ignores control msgs and background How about HDFS reads and writes – Only focus on intermediate data Sub optimal control loop Benefits for small jobs?

16. CoFlows : Proactive Approach Modify the applications – Have them directly inform network of intent Application inform network of co-flow – Co-flow: Group of flows bound by app level semantics – Challenges: End-points not known at the beginning of transfer Start times of the different flows not know File-sizes not known but can be estimated

17. Interactions between Coflows 17 Sharing:  Sharing the cluster network among multiple coflows: How to allocate  Reservation  Max-min faireness Prioritization:  Using priorities as weights  Per job/application

18. We Want To… Better schedule the network  Intra-coflow  Inter-coflow Write the communication layer of a new application  Without reinventing the wheel Add unsupported coflows to an application, or Replace an existing coflow implementation  Independent of applications 18

19. Coflow APIs 19 Get+put operations allow you to overcome the limitation of unknown start times. The network determines when to do the transfer. You can call put, without specifying an endpoint. The network determines where to temporarily store. When the receiver calls a get, the network determines when to transfer the file, what rate and which replica.

20. 20 Coflow API Shuffle finishes MapReduce Job finishes create(SHUFFLE)  handle put(handle, id, content) get(handle, id)  content terminate(handle) Driver

21. Summary Applications know a lot about the transfers – We can reactively learn by using logs – Or modify the application to inform us of these things Tricky information to obtain include: – Transfer start time – Transfer end-points CoFlows: proactive – Controls network path, transfer times, and transfer rate FlowComb: reactive – Controls network paths based on app knowledge 21

22. ToDo Need more images from the infobox guys – Maybe improvements and why skethcy – Maybe graphs from flowcomb also Extensive discussion of pro-active versus reactive. Discussion on orchester should also include patterns from co-flow Add IBM talk.