SEDA [37] reactively provisions CPU resources (by varying the number of threads) across multiple service stages within a single server. SEDA emphasizes request admission control to degrade gracefully in overload. A utility can avoid overload by expanding resource slices and recruiting additional servers as traffic increases. SEDA does not address performance isolation for shared servers.
Resource management for cluster utilities.
Several studies use workload profiling to estimate the resource savings of
multiplexing workloads in a shared utility. They focus on the probability of exceeding
performance requirements for various degrees of
CPU overbooking [4,6,34].
Our approach varies the degree of overbooking to adapt to
load changes, but our current models consider only average-case service quality
within each interval. The target utilization parameters (
)
allow an external policy to control the ``headroom'' to handle predicted load
bursts with low probability of violating SLA targets.
There is a growing body of work on adaptive resource management for cluster utilities under time-varying load. Our work builds on Muse [12], which uses a feedback-controlled policy manager and redirecting switch to partition cluster resources; [39] describes a similar system that includes a priority-based admission control scheme to limit load. Levy [21] presents an enhanced framework for dynamic SOAP Web services, based on a flexible combining function to optimize configurable class-specific and cluster-specific objectives. Liu [22] proposes provisioning policies (SLAP) for e-commerce traffic in a Web utility, and focuses on maximizing SLA profits. These systems incorporate analytical models of CPU behavior; MBRP extends them to provision for multiple resources including cluster memory and storage throughput.
Several of these policy-based systems rely on resource control mechanisms--such as Resource Containers [9] or VMware ESX [36]-- to allow performance-isolated server consolidation. Cluster Reserves [7] extends a server resource control mechanism (Resource Containers) to a cluster. These mechanisms are designed to enforce a provisioning policy, but they do not define the policy. Cluster Reserves adjusts shares on individual servers to bound aggregate usage for each hosted service. It is useful for utilities that do not manage request traffic within the cluster. In contrast, our approach uses server-level (rather than cluster-level) resource control mechanisms in concert with redirecting switches.
The more recent Neptune [32] work proposes an alternative to provisioning (or partitioning) cluster resources. Neptune maximizes an abstract SLA yield metric, similarly to other utility resource managers [12,21,22]. Like SLAP, each server schedules requests locally to maximize per-request yield. Neptune has no explicit provisioning; it distributes requests for all services evenly across all servers, and relies on local schedulers to maximize global yield. While this approach is simple and fully decentralized, it precludes partitioning the cluster for software heterogeneity [5,25], memory locality [14,26], replica control [17,31], or performance-aware storage placement [3].
Virtual Services [29] proposes managing services with multiple tiers on different servers. This paper shows how MBRP coordinates provisioning of server and storage tiers; we believe that MBRP can extend to multi-tier services.
Memory/storage management. Kelly [20] proposes a Web proxy cache replacement scheme that considers the origin server's value in its eviction choices. Storage-Aware Caching [16] develops kernel-based I/O caching heuristics that favor blocks from slower storage units. MBRP shares the goal of a differentiated caching service, but approaches it by provisioning memory shares based on a predictive model of cache behavior, rather than augmenting the replacement policy. MBRP supports a wide range of system goals in a simple and direct way, but its benefits are limited to applications for which the system has accurate models.
Several systems use application-specific knowledge to manage I/O caching and prefetching. Patterson et. al. [28] uses application knowledge and a cost/benefit algorithm to manage resources in a shared I/O cache and storage system. Faloutsos [15] uses knowledge of database access patterns to predict the marginal benefit of memory to reduce I/O demands for database queries.
Hippodrome [3] automatically assigns workload components to storage units in a utility data center. Our work is complementary and uses a closely related approach. Hippodrome is model-based in that it incorporates detailed models of storage system performance, but it has no model of the applications themselves; in particular, it cannot predict the effect of its choices on application service quality. Hippodrome employs a backtracking assignment planner (Ergastulum); we believe that a similar approach could handle server assignment in a utility data center using our MBRP primitives.