PASE: Synthesizing Existing Transport Strategies for Near-Optimal Data Center Transport

Scheduling Mix-flows in Commodity Datacenters with Karuna

“…Information passing: For type 1 and 2 flows, Karuna needs to get the flow information (i.e., sizes and deadlines) to enforce flow scheduling. Such information is also required by previous works [4,22,30,38,39]. Flow information can be obtained by patching applications in user space.…”

“…Similar to [40], we modify timestamp to microsecond granularity. Switch configuration: Karuna only requires ECN and strict priority queueing, both of which are available in existing commodity switches [4,5,30]. We enforce strict priority queueing at the switches and classify packets based on the DSCP field.…”

“…Type 2 flows are still scheduled using SJF. This is the strategy adopted in prior works [4,22,30]. In Figure 2, we observe that the tail latency of type 2 flows increases with the percentage of deadline traffic.…”

“…Further, the sizes of these flows are usually known before transmission. If the traffic is only of this type, SJF-based proposals (e.g., PASE [30] and pFabric with priority given to flows with smaller sizes, or PDQ and pFabric using remaining size as flow criticality) are the best known schemes to minimize average FCT (AFCT). Type 3: Flows without deadlines or known sizes: A number of applications are unable to provide size/deadline information at the start of their flows, e.g.…”

“…For example, PDQ [22] and PIAS [5] do not consider the mix-flow scenario, while pFabric [4] simply prioritizes deadline flows over non-deadline traffic, which is problematic ( §2). Furthermore, many of these solutions [4,22,30,39] require non-trivial switch modifications or complex arbitration control planes, making them hard to deploy in practice.…”