KVM IO performance test data

at present , Most mainstream shared cloud providers adopt hypervisor still XEN, therefore , Many people will generally think that xen Than kvm Better . Personally think that , This is a mistake . foreign amazon And Alibaba cloud in China , The reason why Huawei cloud chose xen As hypervisor, Personally, I think they are only engaged in hypervisor Earlier research on , The best hypervisor Namely xen, Over the years, a lot of manpower and material resources have been invested , A certain amount of knowledge has been accumulated , That's why I always use xen nothing more , That is to say, the decisive factor is not , Technically speaking ,xen Than kvm Better . contrary , I think instead ,kvm In technology, we are on the xen In front of .

1. KVM IO Performance testing

Let's introduce ,kvm stay IO One aspect new feature - virtio-blk-data-plane, First, let's take a look at the performance test results

（1） Test environment configuration parameters

（2） test result

Summary of test results ：

（1） stay block size = 4K Under the circumstances , Adopted virtio-blk-data-plane Of KVM The performance of is the best , Reached 1577684 IOPS, yes sphere 5.1 Of 1.5 times , yes hyper-v Of 4 times , yes KVM Default adopted io drive virt-blk Of 10 times

（2） stay block size from 512B To 8K Under the circumstances , Adopted virtio-blk-data-plane Of KVM The performance is also the best .

2. virtio-blk-data-plane The basic principle of

virtio-blk-data-plane It's in the original virt-blk On the basis of , By giving each io device Create a special thread to take charge of the device io operation , The purpose is to allow different devices IO It can be carried out concurrently . Because each io The equipment io Are handled by a separate thread , One side , Reduced multiple io Synchronization between devices , On the other hand , Less io Operation and qemu Synchronization between main processes . in addition ,io Thread uses io eventfd/irqfd Mechanism , bring io Deal with guest os Implementation decoupling of , Make the most of it host os Of AIO, therefore , Greatly improved guest os Of io performance .

virtio-blk-data-plane The architecture of is shown in the figure below .

because virtio-blk-data-plane stay redhat 6.4 Of tecnology preview, therefore , There are still many restrictions , Mainly in storage ：

（1） Support only raw Image of format

（2）KVM Some of the newer features, It's not available yet virtio-blk-data-plane, such as storage migration, disk hot unplug, io throtting, image streaming and driver mirroring.