Project Insights
To illustrate the benefits of SLAB, let us consider one example scenario in the testbed setup. We consider a 2-tier Clos topology with 4 leaf (or top-of-rack) switches, 4 spine switches, and 4 hosts per-rack, as shown in figure below. All links have a capacity of 10Gbps. We consider two well-known DC workloads: web search and data mining. To generate traffic, we used left-to-right communication patterns. With left-to-right, 8 hosts in the left subtree generate traffic towards 8 hosts in the right subtree. This is a common scenario in user-facing web services where the front-end servers and the back-end storage reside in separate racks. On these settings, we test SLAB, Equal-Cost Multipath (ECMP), a widely deployed scheme in DC, along with Packet Spray (PS), and Weighted Flow Cell Spray.
We found that SLAB performs much better than PS, WFCS and ECMP, both in case of web search and data mining workloads. SLAB creates symmetric virtual topologies (VTs) from asymmetric links. Elephant flows are forwarded through a VT that offers maximum capacity. Whereas, mice flows are probabilistically forwarded to any of the VTs based on their relative capacities. ECMP, on the other hand, maps flows to failed link and fully functional links with equal probability because ECMP is unaware of partial failures in the network. ECMP also suffers from HASH collisions. That is why ECMP does much worse than SLAB. PS sprays packet across all links, and thus its performance is contingent upon the slowest link. For this reason, PS does worst amongst all. The performace of schemes has been evaluated on the basis of flow completion time (AFCT) taken by flows that traversed the network. The schemes have been tested on the network load ranging from 20% to 80%. Certainly, scheme that gives smallest FCTs, i.e., SLAB, performs better than others. Results have been shown in the figures below.
