210 likes | 299 Views
Hierarchy-Aware Distributed Overlays in Data Centers using DC2. Karthik Nagaraj , Hitesh Khandelwal , Charles Killian, Ramana Rao Kompella Computer Science, Purdue University. Today’s Data Center Networks. Up to 240:1 Over-subscription. Core. Aggregation. ... . . ... . . ToR.
E N D
Hierarchy-Aware Distributed Overlays in Data Centers using DC2 Karthik Nagaraj, Hitesh Khandelwal, Charles Killian, RamanaRaoKompella Computer Science, Purdue University
Today’s Data Center Networks Up to 240:1 Over-subscription Core Aggregation ... ... ... ... ToR #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl #incl Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Today’s Data Center Networks Internet • Canonical applications • Group communication systems: Multicast • (Social networks) • Distributed key-value stores • (E-commerce) ... ... ... ... ... ... ... ... ... ... ... ... • Wide-area links • Latency high • Bandwidth low Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Latency Sensitive Routing Scribe overlay multicast (latency-aware) Latency ≠ Location Nearby subscribers unknown No low latency next hop in overlay ... ... ... ... Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Latency Sensitive Routing Ideal overlay tree ... ... ... ... Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Overview • DC2 design • Data Center aware Distributed Communication • Two distributed applications: • DC2-Multicast • DC2-Store • Prototype evaluation • Conclusion Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2 Design • Goals • Limit traffic within close proximity - hierarchy • Provide scalable overlay properties • Components: • Location identifiers • Cluster Coordinators • Overlay routing Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Rendezvous Tree Routing • Hash topic name group key • Route to group key on overlay • Pastry, Chord Rendezvous node Key: SHA160/122a4a9fd2… Intermediate nodes automatically added Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Location Identifiers • Hierarchical Internet Datacenter ID Rack ID Host ID • How we get them • How we use them • Location bit sequence • Prefix node’s hash ... ... ... ... ... ... ... ... ... ... ... ... Rack ID Host ID Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Cluster Coordinators Rendezvous node • Random rendezvous points • Each cluster 0x122a… 0x522a… Rack cluster 0x512a… Location-aware routing Datacenter cluster Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2-Multicast • Location-aware group multicast • Two key ideas: • Contain content distribution within cluster • Avoid forwarder nodes • Forwarder nodes forced into tree • Stress on nodes high Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2-Multicast Rendezvous node • Coordinators provide group lookup • Basic random tree protocol Rack cluster • Rack: 20-40 machines • Datacenter: 100s of racks Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2-Store • Reverse path caching Rendezvous node Traditional Caching DHT Cache Rack coordinator cache Increasing distance Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Evaluation • Deployable Prototype implementation • Using Mace • LOC: 1.1k Mace C++; 22k generated C++ • ModelNet emulated data center topology • 700 nodes • 15 physical machines • Intermediate multicast solution • DC2-Scribe Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2-Multicast Performance • Workload: zipf-ian group sizes, 10 pkts/sec • Latency: Application perceived performance Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2-Multicast Performance • Link Stress: Datacenter network congestion Hierarchy-Aware Distributed Overlays in Data Centers using DC2
DC2-Store Performance • Key-Value lookup latency Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Conclusions • Data center networks over-subscribed • DC2: Explicitly, transparently leverage location • Framework for Distributed communication • Scalable • High throughput • Low latency Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Backup Hierarchy-Aware Distributed Overlays in Data Centers using DC2
Oversubscription • DC2 is a software approach to alleviate condition • Previous hardware approaches • Fat-tree [Alfares08], VL2 [Greenberg09] • Requires changes to existing topologies • Not yet widely used in data centers • Across DCs still an issue Hierarchy-Aware Distributed Overlays in Data Centers using DC2
RandTree protocol • Simple tree management • Efficient on small groups • Assumes lookup service Root Hierarchy-Aware Distributed Overlays in Data Centers using DC2