在雲端系統或者大數據分析平台中，Message Broker扮演著各系統元件訊息交換的重要角色，一般來說在多個租戶的雲端服務上，每個租戶的資源雖然都應該是被隔離的，但由於Message Broker在平台上依舊是被眾多租戶共享的，所以當某部分租戶提交了運算複雜度較高的任務時，Message Broker資源就會被大量消耗，使的無法即時處理到其他租戶的訊息收送，造成其他租戶的整體服務品質降低。上述造成Message Broker資源被大量消耗的租戶被稱為Bad Neighbor，為了解決Bad Neighbor會大量占用Message Broker資源的情形，我們提出了一種適用於多負載平衡的動態頻寬限速分配機制，在有多負載平衡的架構下，使用Message Broker plugin與其提供的API來找出Bad Neighbor，並搭配Linux traffic control在負載平衡伺服器上來對Bad Neighbor做限速。對於各個負載平衡伺服器上的限速頻寬比例分配，由於考量到用戶連線用量不一定會很平均的分布在兩台負載平衡器上，所以我們採用偵測當前負載平衡伺服器上的網路流量來做動態的頻寬限速分配，使得Bad Neighbor即使在被限速後，比起單純使用平均限速頻寬分配的方式，也能夠較有效率得完成提交任務。;In cloud system or big data analysis platform, Message Broker plays an important role in the message exchange of various system components. Generally speaking, the resources of each tenant should be isolated in multi-tenant cloud. However, Message Broker is still shared by many tenants on the same platform. When a certain tenants submit tasks with high computational complexity, the Message Broker resource will be consumed so heavily that the message delivery of other tenants cannot be processed immediately, resulting in lowering overall service quality of other tenants. The above-mentioned tenants who cause the Message Broker resources to be consumed heavily are called Bad Neighbor. In order to solve the situation that Bad Neighbor will occupy a large amount of Message Broker resources, we propose a dynamic bandwidth rate limiting allocation mechanism, which is suitable for multiple load balance. Under the load-balanced architecture, the Message Broker plugin and its API are used to find Bad Neighbor. Besides, Linux traffic control on the load balancing server is utilized to limit the speed of Bad Neighbor. For limiting rate of bandwidth allocation on each load balancing server, we use the network traffic flow on the current load balancing server since the user connection usage amount is not necessarily evenly distributed across the two load balancers. The network traffic flow is used to allocate the dynamic bandwidth speed limitation, which makes the Bad Neighbor able to complete the submission task more efficiently than the average speed-limiting bandwidth allocation method, even after being limited by rate.
