home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.protocols.nfs:2290 comp.protocols.tcp-ip.ibmpc:5165 comp.dcom.lans.ethernet:1875
- Path: sparky!uunet!dtix!darwin.sura.net!zaphod.mps.ohio-state.edu!usc!cs.utexas.edu!torn!watserv2.uwaterloo.ca!watserv1!sail.uwaterloo.ca!eengelke
- From: eengelke@sail.uwaterloo.ca (Erick Engelke)
- Newsgroups: comp.protocols.nfs,comp.protocols.tcp-ip.ibmpc,comp.dcom.lans.ethernet
- Subject: Re: Sun <=> PC Transfer Rate (Summary)
- Message-ID: <BuDFv1.3n6@watserv1.uwaterloo.ca>
- Date: 10 Sep 92 16:45:00 GMT
- References: <1992Sep4.173931.24033@pixel.kodak.com> <karl.23.716089551@empirical.com>
- Sender: news@watserv1.uwaterloo.ca
- Organization: University of Waterloo
- Lines: 75
-
- karl@empirical.com (Karl Auerbach) writes:
- > tgl@ssd.kodak.com (Tom Lathrop) writes:
- >>I was asked to summarize the responses to my query about optimizing
- >>Sun to PC file transfer.
- >
- >And if you use TCP with the new window extension mechanisms you can get
- >much better throughput. However, I don't know any PC code that uses those
- >yet.
-
- The closest I know of is my own FTP which uses up to 32K transmit window.
- I kept the receive window small so that people wouldn't be conjesting
- long haul networks. I know that seems counter productive, but my own
- site is sufferring badly from saturated connections...
-
- >
- >There is also a substantial difference depending on the latency and error
- >rate of the network. If you are going through routers and over the internet
- >you can see a major difference between NFS and TCP file transfer rates.
- >
- >But the #1 absolute most important factor is the speed of the disk.
- >
-
- My PC FTP runs the disk and network concurrently (on large files) so it
- will typically be very much limited to the slower of the source or sink,
- namely the disk. Many of the commercial PC TCP implementations also
- provide this sort of design with varied results.
-
- >I routinely get 2megabit FTP transfers from a Sun Sparc 2 to a wimpy
- >16mhz 386sx box with an 8 bit WD board and packet driver.... as long as the
- >destination is NUL: (Software is PC/TCP v2.1)
- >
- >I've gotten as high as 5megabits/second using a 50Mhz 486DX box with a 16
- >bit WD board using the same software and packet driver.
-
- That is about the limit of the current ISA ethernet hardware and packet
- drivers. 3Com's new (unreleased) cards should up that by 25% to 50%.
- On a different ISA network card we get 850 KB/s, but you really start
- hitting the hardware bus limitations.
-
- >
- >> The CPU is probably not critical,
- >
- >CPU *is* critical. One of the big bottlenecks is checksum calculation.
- >(I do hope that whoever measured the Sun NFS speed did it with checksums
- >turned on -- Sun workstations run with UDP checksums OFF by default.)
- >
-
- Not on a modern system. Times vary greatly, but my desktop 386-33 can
- checksum 24 MB/s, so it's an overhead of about 2%. Variance due to
- network conjestion and clock accuracy far outweigh the effects of
- checksumming when done optimally.
-
- But you raise the point of whether or not it is implemented efficiently,
- that 24 MB/s figure is using 32 bit code. Optimal 16 bit code is much
- slower, and typical (less optimized) 16 bit code is slower still.
-
- >
- >>software side, a couple of people said that if a packet driver is used,
- >>which one it is makes a big difference.
- >
- >I've sent over 2500 packets per second through packet drivers. And that
- >was on a 12mhz 286. Of course, that was a special test designed to
- >load the network.
-
- For SMC cards, the SMC packet driver is much faster than the CRYNWR packet
- driver, but not as reliable on fast hardware. If you want to see how
- many packets/s your packet driver can do, ftp to sunee.uwaterloo.ca
- and get pub/wattcp/speed.zip. Please don't use it on a production network,
- use a test segment instead.
-
- Erick
- --
- Erick Engelke Engineering Computing
- University of Waterloo
- Waterloo TCP Architect erick@development.watstar.uwaterloo.ca
-