Re: Is nfs reliable?

From: Tim Gorman <tim_at_evdbt.com>
Date: Fri, 3 Jun 2016 09:50:59 -0600
Message-ID: <71b3cd50-d934-8ffb-e880-d44feb8f2785_at_evdbt.com>



Anything using UDP is inherently unreliable, because speed, not reliability, is the primary design point. As concisely stated on Wikipedia...



/UDP (User Datagram Protocol) is an alternative communications protocol to Transmission Control Protocol (TCP) used primarily for establishing low-latency and *loss tolerating connections* between applications on the Internet./

The phrase "loss-tolerating" is another way of saying "unreliable"; depends on your perspective. Anyone who deploys NFS over UDP today, even using the most-recent versions of NFS, will still describe it as "unreliable" because of the design of the underlying protocol.

Conversely, anyone who deploys NFS over TCP at any time cannot describe it as "unreliable", but they might be able to describe it as "slow" in relative terms to UDP, because of the design of the protocol.

Not that NFS over TCP is slow. Ethernet is progressing much faster than fiber-channel, where today 100Gb/s is becoming commercially available while FC is hoping to get to 32 Gb/s. There's lots to be found from both perspectives when googling "ethernet vs fiber channel", one of which is
"http://www.mellanox.com/blog/2015/12/top-7-reasons-why-fibre-channel-is-doomed/". There's an excellent chance that, in the near future, conventional wisdom will have NFS as fast as direct attach.

Anyway, all that aside, back to the topic of UDP, I had the pleasure to tweet a nerdy cartoon joke just a few days ago...

https://twitter.com/timothyjgorman/status/737691008436043777

On 6/3/16 08:53, Michael Cunningham wrote:
> Thank you Tim,
>
> I have used nfs for storage since at least 2000 and always found it
> the easiest to configure. With 10g ethernet we got awesome performance.
>
> Since I have never had a problem with nfs I just had to ask Oracle-L
> if anyone had ever heard of nfs being unreliable.
>
> Michael
>
> On Fri, Jun 3, 2016 at 6:39 AM, Tim Gorman <tim_at_evdbt.com
> <mailto:tim_at_evdbt.com>> wrote:
>
> UDP is an unreliable protocol, whereas TCP is reliable.
> Initially, the default protocol for NFS was UDP. Since NFSv3, TCP
> has been an option and is now the default with NFSv4.
>
> NFS over TCP is as reliable as locally-attached storage or SAN
> storage.
>
>
>
>
>
> On 6/2/16 23:02, Michael Cunningham wrote:
>> I had someone tell me today that nfs should not be relied on and
>> it should not be used for a shared mount that needed to be
>> reliably available.
>>
>> Has anyone ever hear this before?
>>
>> --
>> Michael Cunningham
>
>
>
>
> --
> Michael Cunningham

--
http://www.freelists.org/webpage/oracle-l
Received on Fri Jun 03 2016 - 17:50:59 CEST

Original text of this message