[comp.protocols.tcp-ip.ibmpc] Distributed File System speed comparisons

ljm@TWG.COM (12/01/90)

>...I don't consider
>it likely that NFS (an open protocol with end-to-end data checksumming and a
>general purpose machine being used as a fileserver) will ever be as fast as
>the DOS-specific filesharing schemes (lightweight transports, no end-to-end
>checksum, fileserver with optimized disk layout).
>

I'm not so sure.  First of all, we've found that TCP/IP's sophisticated
window management (allowing delayed acknoledgements), average round trip
time
estimation and exponential backoff retransmission algorithms (allowing
relatively fast initial retransmissions on a direct connection), and
dynamic
MTU estimation (allowing full size packets on direct connections) allows
NetBIOS over TCP/IP to ourperform the 'lightweight' NetBEUI protocol stack
by about 20% in PCMAGLAN benchmarks (using LAN Manager).  This disparity
grows much larger if bulk data transfer is the benchmark (the pipe is
always full and every packet wants to be maximum MTU).

Second, in the NFS case, it isn't as if the processing overhead of either
UDP or IP is going to be all that noticeable.  Moreover, using TCPish
retransmission algorithms and 8K or better read/writes yields very good
NFS client performance.

Third, the new special purpose NFS servers (e.g. Auspex) have load
performance
characteristics to be measured in terms of ethernets rather than hosts. 
And, (yes, I know this borders on the religious) given a *reasonable*
implementation
a stateless file server should yield better overall performance than a
connection oriented server.

enjoy,
leo j mclaughlin iii
The Wollongong Group
ljm@twg.com

backman@vaxeline.COM (Larry Backman) (12/03/90)

In article <9011281711.aa22418@Mercury.TWG.COM> ljm@TWG.COM writes:
>>...I don't consider
>>it likely that NFS (an open protocol with end-to-end data checksumming and a
>>general purpose machine being used as a fileserver) will ever be as fast as
>>the DOS-specific filesharing schemes (lightweight transports, no end-to-end
>>checksum, fileserver with optimized disk layout).
>I'm not so sure.  First of all, we've found that TCP/IP's sophisticated
>window management (allowing delayed acknoledgements), average round trip
>....
>....
>Third, the new special purpose NFS servers (e.g. Auspex) have load
>performance
>characteristics to be measured in terms of ethernets rather than hosts. 
>And, (yes, I know this borders on the religious) given a *reasonable*
>implementation
>a stateless file server should yield better overall performance than a
>connection oriented server.
>

I would agree with Leo; I've spent 2 years as a NETBEUI user/developer
and wasn't impressed with its *real life * performance.  Yes, I too
have seen, and even produced the benchmarks that show that 2 NETBEUI
workstations with fast enough cards can saturate an Ethernet; I also
know how the performance was measured. Private net, no bridges, no
routers, and certainly, no other active traffic!

Netbeui gets ugly quickly on noisy nets; its backoff didn't seem as
robust as TCP's; Netware's backoff is even worse!

Likewise, neither NETBEUI or Netware are anywhere near as scalable
as TCP.  Its amusing to watch Netware or NETBEUI on a complex router
based network.

As to Leo's final point, that faster NFS servers are coming ion the
the market; thatnk goodness, its about time.  I know to what lengths,
almost maniacal at times, the designers of both Netware & LAN Manager
have gone to optimize their file server performance.  The architects &
developers of both servers talk in terms of ...300 instructions from
network packet to disk queue...only xxx microsecond disk I/O latency...
and so on.

Eventually someone is going to do the same with an NFS server, and you
are going to see some more equivilent performance through an NFS net.

Another point; Novell's Netware Core Protocol (NCP) have over 120
RPC's last time I looked; Micrsoft's Server message Block (SMB)
protocol has over 60 RPC's with an entire protocol (Named Pipes) overlaid
on top of a single RPC.  NFS on the other hand has roughly 20 RPC's.
Which server do you think is going to give better performance once
platforms &  servers start to become more equal?


				Larry Backman
				backman@ftp.com