Presentation is loading. Please wait.

Presentation is loading. Please wait.

Multi-Core Parallel Routing

Similar presentations


Presentation on theme: "Multi-Core Parallel Routing"— Presentation transcript:

1 Multi-Core Parallel Routing
05/01/ NetSciCom Multi-Core Parallel Routing Multi-Core Parallel Routing - Ahmet Soran*, Mehmet Hadi Gunes*, Murat Yuksel⁺ *University of Nevada Reno +University of Central Florida

2 Shortest-Path Routing
Internet routing: shortest-path Greedy (shortest path) Single path Advantage: easy to compute (Dijsktra in OSPF) 05/01/ NetSciCom Multi-Core Parallel Routing - Maximum Multi-Commodity Flow Problem: Multiple commodities between different node pairs Pass the maximum flow through a network No restriction on the path characteristics No restriction on the flow rates Drawback: NP-hard 5 units 2 units How to maximize throughput for more than one pair? Leverage Multi-Core CPUs for Parallelism?

3 Multi-Path Routing - Focus
How to maximize sent data between pairs? Multipath routing (heuristics) Requires non-shortest path Mostly done via prior centralized calculations Requires significant updates to the routers 05/01/ NetSciCom Leverage Multi-Core CPUs 2 units (s-o-q-t) Multi-Core Parallel Routing - 1 units (s-o-q-r-t) MaxFlow – Multi Commodity Flow Problem: Pass the maximum flow through a network No restriction on the path characteristics No restriction on the flow rates Drawback: NP-hard 2 units (s-p-r-t) 5 units

4 Motivation Multi-core routers are available …
Routers already know how to calculate shortest-path How can we use them for improving routing protocols? 05/01/ NetSciCom Split topology into multiple substrates Assign each substrate to a different core Calculate shortest-paths in each substrate Multi-Core Parallel Routing - Parallelization End-to-end data flow optimization Comprehensive Solution Fast solutions Adaptive solutions Improvable solutions

5 How to generate a new substrate efficiently?
05/01/ NetSciCom How to generate a new substrate efficiently? Multi-Core Parallel Routing - Max Flow = N power 3 Multi commodity problem -> np hard

6 Substrates – Core Solution
Total number of substrate sets can be produced : 2 𝐸 𝑆 How to produce new substrate? 05/01/ NetSciCom Multi-Core Parallel Routing -

7 Edge Based Removal Most Used Edge First 05/01/2017 - NetSciCom
Multi-Core Parallel Routing -

8 How to Generate Substrates
Graph-Based: Which Node to Remove? Heuristic Step: Remove the node most central to the network Betweenness Centrality # of SPs that pass through Flow-Based: Which Edge to Remove? Heuristic Step: Remove the most utilized edge Can capture dynamism in traffic and network 05/01/ NetSciCom Multi-Core Parallel Routing -

9 Substrates: Heuristics
Maximum aggregate throughput Minimum process time Goals Keep a Substrate 0 Iteratively remove parts of S0 Sanity Node removing Edge removing Approaches the most diverse non-overlapping shortest paths Intuition 05/01/ NetSciCom Multi-Core Parallel Routing -

10 Substrates: Which Elements to Remove?
Network Centrality Metrics Betweenness Centrality # of SPs that pass through Closeness Centrality ∑ length of SP from the node Degree Centrality Node degree Edge Betweenness Centrality Page Rank Centrality EigenVector Centrality 05/01/ NetSciCom Calculate Scores Choose elements from ordered list Remove elements Create the substrate Multi-Core Parallel Routing - Heuristic Step: Remove the node most central to the network. Independent – subgraphs are generated from the entire actual topology Cumulative – subgraphs are generated from the previous one sequentially. Subgraphs get smaller with each generation

11 Topologies - Rocketfuel
05/01/ NetSciCom 1221 (Telstra) 1239 (SprintLink) 1755 (Ebone) 3257 (Tiscali) Multi-Core Parallel Routing - Telstra Sprintlink Ebone Tiscali Exodus AboveNet 3967 (Exodus) 6461 (AboveNet)

12 Simulation Flow Set Capacity sharing Highest Flow Random All pair
𝐹𝑙𝑜𝑤 𝐷𝑒𝑚𝑎𝑛𝑑= 𝑝𝑜𝑝1 ∗𝑝𝑜𝑝2 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 2 Capacity sharing Max-min allocation Proportionally with flow demand Highest Flow Potential maximum flow Random Lower bound 05/01/ NetSciCom Multi-Core Parallel Routing - It is not a packet simulator, static analysis .. HF removes the most utilized edges in each step HF needs to be recalculated HF is dynamic but computationally intensive Flow demand is product of populations divided by the square of the geo-distance between two routers

13 Centrality Analysis Multi-Core Parallel Routing -
05/01/ NetSciCom Multi-Core Parallel Routing - In the figure, x-axis has two different indicators. First line on the top shows the removal percentage of network nodes/links (which is incremented between 2% and 16% by 2). Second line gives the number of cores (i.e., 2, 4, 8, 16, 32 cores) that will generate subgraphs. We observe that MCPR heuristics improve with the number of cores node centrality heuristics perform slightly worse than random network generation, centrality heuristics overall yield better results than random removals edge centrality heuristics are better with high number of cores and independent removal percentage within 8% to 10%.

14 Centrality Analysis Multi-Core Parallel Routing -
05/01/ NetSciCom Multi-Core Parallel Routing - the independent method removes fewer elements in generating the next subgraph. we observe that the cumulative method performs worse with higher node/edge removal rates they are not able to yield viable e2e paths HF removes hot spot(s) from the subsequent subgraphs

15 Number of Core Analysis – Independent (%8 Removal)
05/01/ NetSciCom Multi-Core Parallel Routing - Figure 4 presents the effect of the number of cores in the networks with %8 independent edge removal. As observed in both the Telstra and overall average results, the speedup performance improves with the number of cores

16 Substrate Changes Multi-Core Parallel Routing - 05/01/2017 - NetSciCom
node centrality heuristics disrupt the network in subgraphs more than edge centrality and highest flow heuristics We observe that small removal percentages have low effects on subgraph characteristics

17 Substrate Changes Multi-Core Parallel Routing - 05/01/2017 - NetSciCom
Cumulative subgraph generation significantly changes graph characteristic and removes connectivity in the graph

18 Substrate Changes Multi-Core Parallel Routing - 05/01/2017 - NetSciCom
Average node degree, however, is not effected much by the edge removal heuristics. In node centrality heuristics, first subgraph removes the most centralized nodes, which causes significant decrease in the average and maximum node degree. After a point, MCPR start to remove periphery nodes from the network and this increases the average node degree.

19 Substrate Changes Multi-Core Parallel Routing - 05/01/2017 - NetSciCom
Note that the removal percentage should be less than 100/coreNumber as after this point, there are no more nodes to remove in the subgraphs

20 Summary A novel divide-and-conquer approach:
Divide: split topology into multiple virtual substrates Conquer: use shortest-path on each substrate The overall problem is transformed/mitigated into the divide phase Significant improvements via using multi-cores Up to 4 times speed up in aggregate throughput Particularly applicable to data center networking 05/01/ NetSciCom Multi-Core Parallel Routing -


Download ppt "Multi-Core Parallel Routing"

Similar presentations


Ads by Google