Network (SSH) performance on ESXi 5

[dcoppee]

There is a similar discussion (Network/Raid performance on ESXi5) here but as it is marked 'answered' I cannot add any more stuff

Anyway, the answer (set useDNS to no) does not change the problem.

As mentioned in the other discussion, for some reason the speed of a file copy via ssh (e.g between two esxi hosts) does not exceed 11MB/s

Copy speed via virtual machines (also via SSH) on these hosts result in much higher speed

So, whats holding the host access ssh connection back?

I use a shared Dell iDrac connection (server is Dell R610). I wonder if this has impact

regards

Daniel 

[Mauro Bonder]

Discussion moved from VI: VMware ESXi™ 3.5 to VMware ESXi 5

the speed of this host vmware are configured correct ?

[Slingsh0t]

Hard to say without knowing the network setup in your environment.  Are the management nics on the hosts all on the same flat layer 2 network?  Any traffic shaping configured on the host ports?  Can you see the network topping out at 11MB/s in the performance tab of the host with the management NIC selected?

[Freaky201110141]

Not an expert on this, far from it. But I can tell you 2 things that might help.

a) It uses dropbear for SSH. Whilst nice and small, it's very VERY inefficient at encryption (shitty algorythm I guess). Just look at the CPU when doing scp's with dropbear from/to openssh. The box using openssh will hardly use any CPU at all, whilst the box with dropbear can easily max the CPU

b) the console runs in a resource constrained VM (it's a special VM)

AFAIK, the performance you see comes from a combination of the 2 of these. And in all the years I've searched (not very extensive tho'), this has never been solved (quite honestly, from a technical point, I doubt you can solve it). Perhaps if you can find a way to change the encryption used by dropbear to something less CPU intensive that you can get a boost.

HTH

[dcoppee]

Hello,

thanks all already for these Ideas.

What is still strange: copy from old server with ESXi4.1 to old server with ESXi4.1 gives more or less 20 to 25MB/s (limited by the speed of the disks I suppose)

Copy from one of these to the new (fast disks, new strong 2x 4-core, 24GB RAM etc etc) server with ESXi5.0 gives only 11MB/s

So there must be something else that is new to this version of ESXi

The server is a Dell R610, fully hardware compatible

Regards

Daniel

[iwienand]

Are you copying to local disk or to a SAN?

Anyway, the bottleneck you're seeing is likely VMFS.  Being a distributed file-system meta-data updates are very costly, and it is not designed for the way that scp copies a file by constantly growing it.  Unsurprisingly, VMFS is designed for distributed access to large, preallocated blocks; i.e. virtual machine disk images :smileyhappy:

Unfortunately I don't think there's much you can do about this right now.  However we have heard this request, and we are working with the file-system team to make some ways that scp can indicate its intent so we can optimise performance in future releases. 

[dcoppee]

Hi,

Thanks for clarification about VMFS

In the case of the tests I always used local disks

I just now discovered that if I use VM_Explorer to copy files between hosts, it goes much faster that pure scp.

Same network, same disks, same servers. But with an installed 'agent' (in expert mode).

I'm sure many small companies have to use standard (free) tools like scp to move a virtual server from one host to another instead of using the VCenter software (with Vmotion). Downtime is acceptable when budget is small but not if it takes hours to do this copy/move.

Regards

Daniel Coppee