基于集群虚拟化的高精度快照的设计与实现

Design and implementation of a cluster virtualization based fine snapshot system

提出了一种逻辑卷级别的高精度快照技术,能够有效利用存储资源,为目标逻辑卷生成快照逻辑卷,精确保留目标卷的历史数据,提高系统可靠性。设计了checkpoint和finesnap2种卷级别快照结构,通过结合使用两者,建立checkpoint与finesnap的混合链表,在节省存储资源的同时,维持好的系统性能;通过采用在内存中缓存finesnap位图的策略,加速finesnap中数据的查找,同时将缓存消耗的内存数量控制在合理范围内,实现了一个原型系统LV-Fine。实验证明,对于目标逻辑卷的读写,当快照卷数量为16且finesnap与checkpoint的比例为9:1时,LV-Fine的性能较著名集群虚拟化系统LVM2提高了133%;对于历史数据的空间占用,在相同的快照生成频率finesnap与checkpoint的比例为8:1,且trace播放10h的前提下,LV-Fine的占用量仅为LVM2的37%。

A fine snapshot technology in volume level was presented which could be used to keep old data flexibly and frequently, to recovery from deletion, to limit exposure to data lose for logic volumes. First, two kinds of logic volumes namely, checkpoint and finesnap were designed and used as containers of snapshot data. Linking checkpoint volumes and finesnap volumes together by the order of time could make up of a hybrid snapshot chain, with which both good performance and high efficiency on storage usage were available. Second, a bitmap in each finesnap volume to accelerate data searching and reduce the memory exhausted was incorporated. Evaluation results from representative experiments demonstrate that the prototype system, named LV-Fine, has the ability to provide good performance, and to introduce less storage overhead than LVM2, the famous cluster virtualization system.