Many systems rely on distributed caches with thousands of nodes to improve response times and off-load underlying systems. Large-scale caching presents challenges in terms of resource utilization, load balancing, robustness and flexibility of deployment. In this paper, we propose a novel distributed caching method based on dynamic IP address assignment. Keys are mapped to a large IP address space statically and each node is dynamically assigned multiple IP addresses. As a result, we have a system with minimal need for central coordination, while eliminating the single point of failure in competitive solutions. We evaluate our system in our datacenter and show that our approach localizes the effect of load-balancing to only loaded cache servers, while leaving cache clients unaffected and also providing for finely-granular rebalancing.
Ultra-fast load balancing of distributed key-value stores through network-assisted lookups / De Cesaris, Davide; Katrinis, Kostas; Kotoulas, Spyros; Corradi, Antonio. - STAMPA. - 8632:(2014), pp. 294-305. (Intervento presentato al convegno 20th International Conference on Parallel Processing, Euro-Par 2014 tenutosi a Porto, prt nel 2014) [10.1007/978-3-319-09873-9-25].
Ultra-fast load balancing of distributed key-value stores through network-assisted lookups
CORRADI, ANTONIO
2014
Abstract
Many systems rely on distributed caches with thousands of nodes to improve response times and off-load underlying systems. Large-scale caching presents challenges in terms of resource utilization, load balancing, robustness and flexibility of deployment. In this paper, we propose a novel distributed caching method based on dynamic IP address assignment. Keys are mapped to a large IP address space statically and each node is dynamically assigned multiple IP addresses. As a result, we have a system with minimal need for central coordination, while eliminating the single point of failure in competitive solutions. We evaluate our system in our datacenter and show that our approach localizes the effect of load-balancing to only loaded cache servers, while leaving cache clients unaffected and also providing for finely-granular rebalancing.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.