高级操作系统高级操作系统 (17).pdf

《高级操作系统高级操作系统 (17).pdf》由会员分享，可在线阅读，更多相关《高级操作系统高级操作系统 (17).pdf（13页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、第 4 讲：Optimization of Virtual Machine Monitor第一节：IntroductionIntroductionBefore optimization of VMM Where is the bottlenecks of VMM?How to find these bottlenecks?.Introduction Where is the bottlenecks of VMM?bottlenecks of VMMVCPUVMEMVIOetc.Introduction bottlenecks of VMMComparison between VMM and C

2、ontainerTarget:Docker v.s.KVM/XENBenchmark:Environment andWorkloadTesting&Evaluationfrom:An Updated Performance Comparison of Virtual Machines and Linux Containers,TR of IBM,2014from:My VM is Lighter(and Safer)than your Container and Linux Containers,SOSP17,2017.Introduction Container overview.Intro

3、duction Container overview.Introduction Container overviewWhat is inside a ContainerProgram Binaries/configurationRuntime librariesDependency Products/toolsA Piece of KernelSystem Resources:CPU/MEM/IO/Net/StorageWe are isolating the program and delicatelyproviding its own system resources andruntime

4、 libraries.Introduction bottlenecks of VMM CPU/MEMEnvironmenttwo E5-2665(16 cores with HT),256 GB of RAMDirect 10 Gbps Ethernet link between two Mellanox ConnectX-2 EN NICsUbuntu 13.10,kernel 3.11,Docker 1.0,QEMU 1.5.032vCPU(Power management was disabled).Introduction bottlenecks of VMM CPU/MEMResul

5、tsPXZ:extra TLB pressure of nested paging.Linpack:pin vCPUs pCPU,expose the underlying cache topologySTREAM:bandwidth of mem,cost of handling TLB misses,uses large pages.Introduction bottlenecks of VMM NetNetwork configurationsnuttcp to measure network bandwidthAll three configurations reach 9.3Gbps

6、NAT noticeably increases overheadvhost reduces overheadNIC is the bottleneckTCP bulk transfer efficiency(CPU cycles/byte)Network round-trip latency(s).Introduction bottlenecks of VMM StorageStorage configurations20 TB SSD,two 8 Gbps FibreChannel linksfio with the libaio backend in ODIRECT modeFibre

7、Channel HBA is the bottleneckSequential I/O throughput(MB/s)Random I/O throughput(IOPS).Introduction bottlenecks of VMM Real AppsMySQL configurationsMySQL throughput(transactions/s)Redis performance(requests/s)Average latency(in ms)of operations on Redis.Introduction bottlenecks of VMM SummarySummar

8、yContainers and VMs impose almost no overhead onCPU and memory usage;they only impactI/O and OS interaction.This overhead comes in the form of extra cycles foreach I/O operation,so small I/Os suffer much morethan large ones.Several additional topics worthy of investigation:performance isolation when multiple workloads runon the same server,live resizing of containers andVMs,tradeoffs between scale-up and scale-out,andtradeoffs between live migration and restarting.https:/