Slow NFS copy times

[aleceiffel]

I've setup an nfs share on a windows machine to store iso files as well as VM template files. The windows 2003 R2 SP2 server running Windows services for UNIX 3.5 has 2 network cards, one of which is connected to a 1Gb switch that only has the ESX server interface running VMKernel for NFS attached to it. The NFS share has been setup following the instructions in the following link: .

I can transfer files fine using the datastore browser and vmkfstools but to transfer a 8GB vmdk file takes over an hour. This seems way too slow to be practical.

Does anyone know how I would get useable transfer rates?

[Dave_Mishchenko]

Your post has been moved to the Performance forum

Dave Mishchenko

VMware Communities User Moderator

PS - check the post by Paul Lalonde on this thread to see if you're having the same problem - http://communities.vmware.com/message/944761.

[drummonds]

Hello,

You may have excited a corner case that unfortunately (and necessarily) is very slow. To maintain crash consistency between the source and target, each incoming TCP packet from the NFS store is written directly to the VMFS volume. If you happen to be writing to a local disk or an array with a slow or non-existant write cache, every TCP packet must be written completely to disk before it is ACKed and the next packet can start. ESX does not buffer these packets lest a crash occur and the files become crash inconsistent. There is no workaround for this problem.

However, you may be able to structure your operations to avoid this condition. Consider the following:

1. Do you really need to copy the files? You can directly mount the NFS-based VMDKs and ISOs and not have this problem.

2. Is there a write cache that you can enable? This decreases the write time for each TCP packet and will greatly improve throughput.

3. Can you change the target to be on a different array with a larger write cache?

I know this is a pain. Unfortunately, it's a necessary requirement to maintain crash consistency.

Scott