DCRoute: Speeding up Inter-Datacenter Traffic Allocation while Guaranteeing Deadlines

Transcription

1 DCRoute: Speeding up Inter-Datacenter Traffic Allocation while Guaranteeing Deadlines Mohammad Noormohammadpour, Cauligi S. Raghavendra Ming Hsieh Department of Electrical Engineering University of Southern California Sriram Rao Cloud and Information Services Lab Microsoft 1

2 Datacenters The infrastructure for many Cloud Computing services More of them as more users rely on these services Geographically scattered close to users to reduce latency 2

3 Inter-Datacenter Networks Public or private WAN networks Connect datacenters in different cities / countries / continents Costly to provision and maintain 3

4 Inter-Datacenter Traffic Interactive Directly affects user experience Should be delivered Instantly with strictly higher priority Examples: Web requests Deadline Needs to be delivered prior to a deadline Elastic to bandwidth allocation Considerably large Examples: Geo-replication Background Served in a best-effort manner Examples: Data Warehousing 4

5 Deadline Transfers A deadline transfer is a traffic request shown as R with: A deadline dl(r) or T d (R) Demand vol(r) A source node (datacenter) src(r) A destination node dst R Transfers can originate from or end at any datacenter in the network 5

6 Scheduling Traffic It is important to use inter-datacenter bandwidth efficiently: Over-provisioning is costly Allocating / Scheduling bandwidth can mitigate congestion and improve efficiency We can increase utilization close to 100% A central controller: Receives transfer requests Calculates the allocations for deadline traffic every timeslot The allocations are sent to end-hosts generating requests End-hosts limit their transmission rates as instructed by controller Broker Central Scheduler / Allocator Datacenter 1 Broker Datacenter 2 Broker Datacenter 3 6

7 Previous Work Do not guarantee deadlines: SWAN [sigcomm 13], B4 [sigcomm 13], and BwE [sigcomm 15] aim to maximize utilization and focus on max-min fairness Tempus [sigcomm 14] strives to maximize the minimum proportion of transfers finished before deadlines Amoeba [eurosys 15] guarantees deadlines however: Uses LP modeling to re-calculate allocations upon arrival of every request Does not take into account the effects of packet re-ordering Higher CPU usage, buffering, and latency at the receiving end hosts These get worse at high inter-datacenter rates (gigabits/s) 7

8 Our Traffic Scheduling / Allocation Problem How to schedule large transfers with deadlines over inter-datacenter networks along with interactive and background traffic Guarantee that admitted transfers meet their deadlines Develop a traffic allocation technique that provides: High Utilization: Use WAN resources as much as possible No Reordering: Reduce host resources for putting packets back in order All traffic for a transfer must go on the same path Little Computational Overhead: Large number of transfers have to be processed and allocated 8

9 General Approach Predict and reserve small bandwidth for interactive traffic; treat background as best effort; our problem is allocating large deadline traffic The allocation problem can be modeled as an optimization scenario: Objective: Maximizing a Utility Function (link utilization, fairness, etc.) Constraints: Link Capacities, and Meeting Deadlines If objective and constraints are linear, this forms a linear program (LP) Upon arrival of every request: Create a LP involving all requests with positive residual demands If the LP is feasible, a new allocation is found and request is accepted If the LP is infeasible, the request is rejected We aim to avoid linear programming 9

10 As Late As Possible (ALAP) Scheduling Used in Operations Research Allocate resources when absolutely needed Upon arrival of a request: Start from its deadline moving backwards Sum all residual capacities in timeslots If less than request s demand, cannot accept it Straightforward for a single link NO LP NEEDED! 10

11 Maximizing Utilization Pull back traffic from future timeslots: Upon beginning of a timeslot If current timeslot is not fully utilized Pull from closest timeslots in the future The resulting allocation will still be ALAP A newly arriving request can be processed fast Straightforward for a single link NO LP NEEDED! 11

12 Scheduling in Network Need to allocate resources on all links of a path: Example: Scheduling two transfers from A D and B D; may arrive at different times It may not be possible to allocate ALAP on all edges: Multiple ways to allocate ALAP The optimal solution could be either one We allocate in the order of arrival 12

13 DCRoute Admission Control and Bandwidth Reservation: Allocate(R), R is a transfer request Updating Schedules / Allocations every timeslot: PullBack() PushForward() Dispatching allocations to datacenters: Walk() 13

14 Allocate(R) For every new request submitted: Select a path Schedule R on that path A simple heuristic Based on Dijkstra s shortest path To balance #hops (total bandwidth usage) and distribution of load (equalize load distribution by using longer paths with less load) We will see it in the next two slides Schedule traffic according to ALAP policy To allow fast admission control and scheduling / allocation updates 14

15 Allocate(R) Intuition: A request R is submitted with demand vol(r) Take a path P on inter-datacenter network from src(r) to dst(r) with N hops If R is scheduled on P: Accumulative load over edge e of P for t dl(r) is L e The total cost is noted as the total accumulative load if R is put on P: COST = σ e P (L e + vol R ) This cost accounts for the additional bandwidth N vol R It also accounts for total load of the edges on the selected path Minimizing it will most likely: Avoid very long paths Avoid highly utilized paths To find P, assign cost L e + vol R to each edge of the inter-datacenter graph Run Dijkstra s algorithm to find the path with minimum cost from src(r) to dst(r) 15

16 Example: New request R, vol(r) = 8 Numbers on initial graph (left) represent accumulative load on links prior to dl R Initial Graph: 10 Augmented Graph: 18 Dijkstra: 18 Minimal cost path S 60 T S 68 T S 68 T S T 16

17 PullBack() Maximize bandwidth utilization Scan from next timeslot to the latest deadline Scan all links Move as much allocation as possible to current timeslot Allocations should be updated over all links of every request s path 17

18 PushForward() Make the allocation ALAP Scan from 2 timeslots ahead to the latest deadline Scan all links Move any allocation as close as possible towards their deadlines Allocations should be updated over all links of every request s path 18

19 Example 19

20 Walk() Dispatch rate-allocations to all senders for next timeslot Adjust demands according to how much was transmitted in current timeslot for each request R currently in the system 20

21 Bandwidth Reservation for Future Requests Advance information about a request: A request will arrive some time later Cannot initiate the transfer yet Still need to make sure it makes it by a deadline Can use the same techniques added that: Pull traffic from requests with available data Push forward the traffic that will arrive later 21

22 Evaluation Show Rejection Rate and Allocation Speed We compare DCRoute with Amoeba G-Scale topology, same synthetic traffic distribution as in Amoeba [eurosys 15] We then compare DCRoute with: Global LP: Considers all edges for every request and routes traffic over all possible paths K-Shortest Paths: For every request R, schedules traffic on the K shortest paths from src(r) to dst R Two experiments with with synthetic traffic: 1. Using G-Scale topology Effect of arrival rate 2. Using random topologies with specific network sizes Effect of network size 22

23 % Rejection Rate Speedup Ratio Comparison with Amoeba % Rejection Rate (DCRoute - Amoeba) Speedup Ratio (DCRoute / Amoeba) Request Arrival Rate Request Arrival Rate 23

24 Effect of Request Arrival Rate 24

25 Effect of Network Size 25

26 Conclusions DCRoute: An inter-datacenter WAN routing and scheduling technique Guarantees that transfers make it to their deadlines Avoids reordering by scheduling all packets of a request on the same path Relies on ALAP technique to speed up the scheduling / allocation process: At least 2 orders of magnitude faster than all evaluated schemes Admits at most 4% less traffic compared to all evaluated schemes 26

27 Future Work Evaluation of DCRoute using real inter-datacenter traffic Developing methods to properly re-schedule / re-allocate traffic in the event of link failures to minimize damage How this allocation affects median or tail flow completion times (FCT) Incorporating max-min fairness into DCRoute (Is it possible?) Possibly applying these techniques to small flows with deadlines 27

28 Thank you! Q & A 28