[Qemu-devel] 答复: 答复: question about pe rformance of dataplane

[Zhangleiqiang]

Hi, Stefan:

        Follow your advice, I have finished the benchmarks with multiple vcpu 
(smp) and parallel I/O workloads. 
        The results still show that the performance of disk with dataplane 
enabled did not have advantage over non-dataplane under Random write mode. But 
under the Sequence write mode, the former has obvious advantage.

1. Environment:
        a). Qemu 1.4 master branch
        b). kernel:  3.5.0-2.fc17.x86_64
    c). virtual disks location:  the same local SATA harddisk with ext4 fs
        d). VM start cmd (os: win7/qed, disk1: raw/non-dataplane/10G/NTFS, 
disk2: raw/dataplane/10G/NTFS) :
        e). vcpu: 4

./x86_64-softmmu/qemu-system-x86_64  -enable-kvm -smp 4 -name win7 -M pc-0.15 
-m 1024 -boot c -device virtio-serial-pci,id=virtio-serial0,bus=pci.0,addr=0x5 
-monitor stdio -drive 
file=/home/win7.qed,if=none,format=qed,cache=none,id=drive0 -chardev 
spicevmc,id=charchannel2,name=vdagent -device 
virtserialport,bus=virtio-serial0.0,nr=1,chardev=charchannel2,id=channel2,name=com.redhat.spice.0
 -chardev pty,id=charchannel3 -device 
virtserialport,bus=virtio-serial0.0,nr=2,chardev=charchannel3,id=channel3,name=arbitrary.virtio.serial.port.name
 -usb -device usb-tablet,id=input0 -spice 
port=3007,addr=186.100.8.121,disable-ticketing -vga qxl -global 
qxl.vram_size=67108864 -device AC97,id=sound0,bus=pci.0,addr=0x4 -device 
virtio-blk-pci,drive=drive0,bus=pci.0,addr=0x6,bootindex=1 -drive 
id=drive1,if=none,cache=none,format=raw,file=/home/data.img -device 
virtio-blk-pci,drive=drive1,bus=pci.0,addr=0x7 -drive 
id=drive2,if=none,cache=none,format=raw,file=/home/data2.img,aio=native -device 
virtio-blk-pci,drive=drive2,bus=pci.0,addr=0x8,scsi=off,x-data-plane=on,config-wce=off

2. Testing Tool And Testing Params:
        a). Only IOMeter App Running in VM, and no other IO in Host (Dom0)
        b). 100% Random, 0% Read, 16K data size, 50 outstanding IO;  0% Random, 
25% Read, 16K data size, 50 outstanding IO
        c). Testing of two disks separately, not simultaneously

3. Testing Results:

        RW_mode                         IOPS_dataplane          
IOPS_non_dataplane              MBPS_dataplane          MBPS_non_dataplane
        100% Random     /0% Read                        303.178867              
        300.511928                              4.737170                        
        4.695499
        100% Sequence/25% Read          21748.887189                    
7631.164060                             339.826362                      
119.236938



----------
Leiqzhang

Best Regards

> -----邮件原件-----
> 发件人: Stefan Hajnoczi [mailto:address@hidden
> 发送时间: 2013年4月4日 15:21
> 收件人: Zhangleiqiang
> 抄送: Stefan Hajnoczi; address@hidden; leiqzhang; Haofeng; Luohao
> (brian)
> 主题: Re: 答复: [Qemu-devel] question about performance of dataplane
> 
> On Tue, Apr 02, 2013 at 02:02:54AM +0000, Zhangleiqiang wrote:
> >     I have also finished the perf testing under Fedora 17 using IOZone, and 
> > the
> results also shown that the performance of disk with dataplane enabled did not
> have advantage over non-dataplane.
> 
> virtio-blk data plane is a win for parallel I/O workloads (that means
> iodepth > 1).  The advantage becomes clearer with SMP guests.
> 
> In other words the big advantage is that data plane processes requests
> without blocking the QEMU main loop or vCPU threads.
> 
> If your guest has 1 vCPU and/or your benchmarks only do a single stream
> of I/O requests, then the difference may not be measurable.
> 
> Stefan