"Peter" == Peter D Gray pdg@uow.edu.au writes:
Peter> On Thu, Jan 25, 2018 at 08:44:00AM -0500, John Stoffel wrote:
That's good to hear! The takeaway I have from this is that NFS over UDP is not something you should ever be using.
Mark> I have chased down the person who set the servers up and he was Mark> trying different options to see what gave the best performance Mark> and left udp in the options as while he wasn't sure that it was Mark> any better it hadn't got any worse.
I'm curious about how they did their testing? And what the cutoff was for making changes and whether it was worth keeping or not. All the docs I've read from Netapp and Oracle say to use NFS over tcp, with large read/write block sizes and then some other options in special cases.
In my mind, the advantages of TCP over UDP even for regular NFS traffic make it a no brainer.
Peter> Hmmmm..... disagree.
Peter> In the best of all possible worlds UDP wins. Its fast and you Peter> can overlap multiple reads and write much easier than Peter> TCP. Those guys who invented NFS used it for a reason. If I Peter> wanted raw performance I would use UDP.
They used UDP at the time because computers and networks were *slow* and the TCP overhead was much higher then, esp since they mostly had hubs back then. Under contention, NFS over TCP would slow way down. I would agrue that this is a false economy today when we have 10g networks. *grin*
Peter> However, in lots of cases UDP has problems. Network devices Peter> are often optimized for TCP (firewalls are a prime example) and Peter> as you say packet sizes can be larger with TCP.
Exactly. Chasing a few percent of speed (or even 10%!) by using UDP is not a great idea. Esp since I suspect that NFS over UDP is a much less tested version of the protocol these days.
Peter> I agree that TCP is a better bet in general but I do understand Peter> why people may want to use UDP.
Peter> Its interesting that the new data transfer algorithms seem to Peter> be UDP based. Aspera for example. I wonder if those types of Peter> protocols could make sense for NFS?
If you're willing to do your own congestion control and packet handling, then sure it can make sense. Esp if you're working over a WAN link and you don't mind out of order packets and can handle it better in your own server software. But how many filesystems are doing this? Esp for POSIX compatibility?
John