[vbox-dev] is virtstor being implemented?

[Huihong Luo]

Achim,
 
Thanks for the clarification.
 
I was playing Microsoft Virtual Server 2005 SP1 this morning, and they have SCSI Shunt Driver, a kind of virtstor, and it causes quite a difference when using it, so I was wondering there might be something. Now I know it must be due to their poor implementation on regular disk emulation.
 
VBox team is superb ! The disk performance is impressive !
 
There is one minor thing about fixed size disk creation, it seems to be quite slow. 10G VHD fixed size takes about 30 mins. Is this normal? Microsoft's VHD driver on windows 7 seems to be much faster. I know this isn't very important, as it's not used at runtime.


--- On Mon, 10/18/10, Achim Hasenmüller <achim.hasenmueller at oracle.com> wrote:


From: Achim Hasenmüller <achim.hasenmueller at oracle.com>
Subject: Re: [vbox-dev] is virtstor being implemented?
To: "Huihong Luo" <huisinro at yahoo.com>
Cc: "Alexey Eromenko" <al4321 at gmail.com>, "Liang Suilong" <liangsuilong at gmail.com>, vbox-dev at virtualbox.org
Date: Monday, October 18, 2010, 11:57 PM


What you describe is basically how SATA and SCSI/SAS work :-)


Honestly, we don't see any reason to implement virtstor. We do not expect any performance advantages and whereas a network driver is easy to add/change after install, a disk controller is more complicated and less user friendly.


I would say that there isn't much out there that can compete with VBox when it comes to disk performance. Our performance work is now mostly focused around network as this is clearly where we have room for improvement. We want to continue to improve the E1000 device but for ultimate performance, our efforts are currently focused on virtio. Once we can make full use of receive and send offloading, we should be good.


Achim




On Oct 19, 2010, at 8:14 , Huihong Luo wrote:






yes, it's the speed reason. In theory, virtstor is supposed to be fast. However, since vbox is already doing so well on disk, I am not sure how much further can be improved.
 
In theory, virtstor bypasses many calls, and goes directly to the cheese, i.e., block i/o data transfer. The performance depends how well the mechanism for the transfer between vm and host, which is why I mentioned share folders. (btw, I know the difference between file sysem and block driver, I do that for a living). What I mean is that we can use the same large data transfer mechanism if shared folder's performance is very good. 
 
for example, a user app wants to write 10k data to the disk, it goes to the virtio driver after  a few calls in windows, if the virtstor driver can efficiently (directly) transfer the buffer to the host, perhaps through memory sharing, or pinned physical pages, this should improve the performance.
 
About the guest drivers, windows 7 and vista sp1 introduced some new models that should be even faster than kvm's implementation.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.virtualbox.org/pipermail/vbox-dev/attachments/20101019/61646364/attachment.html>