多承租者的資料中心內即時應用程式之頻寬管理

Abstract

新興的互聯網服務像是搜尋和線上金融服務大都仰賴公有雲或私有的資料中心來處理大規模且不斷增長的大數據. 很多這類的服務具有即時性. 所以在資料中心的應用程式通常需要滿足像是期限(deadlines)的service level objectives (SLOs). 而網路的效能通常是關鍵的因子. 在多承租者(multi-tenant)的資料中心, 共用的資源環境會使承租者得到劇變的網路效能. 然而其中的頻寬管理是複雜且有挑戰性的. 一方面, 承租者要求網路效能需可預測; 另一方面, 服務管理者要求高使用率. 甚至, 承租者之間的公平性也是被要求的重要性質. 這篇論文主要在研究如何改善承租者間還有承租者本身的效率. 而效率的改善必須兼顧可預測性和公平性的限制. 我們先考慮承租者之間的效率問題. 首先, 因為頻寬的隔離可保證預測性, 我們專注在虛擬叢集(virtual cluster)這種資料中心的服務介面. 我們提出一種方法來改善現行預留系統(reservation systems)的低使用率. 再者, 對於預留系統中沒被使用的頻寬, 我們提出一種簡單的市場訂價機制在遵守公平原則下來改善使用率. 最後, 我們在承租者間具有高使用率的頻寬分配下, 研究承租者本身的即時排程的效率問題. 我們觀察到現行排程演算法的缺點, 並提出一個在時變的頻寬分配下仍可達到高效率的排程演算法.

Modern Internet services, like web search, online banking, social networks and various recommendations, rely on cloud or production data centers to process large-scale and ever-growing data distributed in a large number of compute and storage nodes. Many of these services are time-critical implying that the corresponding applications running in the data centers are subject to meet service-level objectives (SLOs), e.g., deadlines. In the distributed environments, network performance is a leading factor for the time-critical applications meeting SLOs. As these data centers are often shared amongst multiple tenants (running diverse applications), the share nature implies that each tenant could receive variable network performance, leading to deviating SLOs. Bandwidth management in multi-tenant datacenters is challenging. At one end, tenants need predictable underlying network performance; at the other end, providers (or operators) need high efficiency, e.g., for paying off immense investments in building datacenters. Furthermore, inter-tenant fairness could be another desirable property in the public cloud for instance. In this dissertation, we study how to improve inter-tenant and intra-tenant efficiency subject to predictability and fairness for time-critical applications in multi-tenant data centers. We first consider inter-tenant efficiency. As bandwidth isolation leads to predictability, we focus on a popular use case where tenants and providers adopt the resource descriptive interface of virtual cluster, which guarantees bandwidth isolation. Based on observations on existing reservation systems, we propose an approach to improve the efficiency of data centers while servicing multiple tenants in an online fashion. However, spare bandwidth resources left in the reservation systems lead to inefficiency. To this end, we further propose a market-clearing mechanism to further improve the inter-tenant efficiency while achieving the desired fairness among tenants. Then, we consider the impact of efficient inter-tenant bandwidth allocation on the intra-tenant efficiency. More specifically, we observe suboptimal application throughput of existing deadline-aware flow schedulers under efficient inter-tenant bandwidth allocation schemes. Accordingly, we propose a new deadline flow scheduling scheme for a tenant running its time-critical applications under efficient inter-tenant bandwidth allocation. For evaluating our proposed approaches, we conduct extensive simulations based on synthetic and real-world traces. The results show that inter-tenant efficiency can be improved over existing reservation systems and tenants' application throughput is increased under efficient inter-tenant bandwidth allocation.