rpbert@phoenix.Princeton.EDU (Raymond Pierrehumbert) (07/23/89)
Can somebody tell me what the bottleneck is on FTP transfer rates for a MacII on ethernet? I am running two MacII's on a subnet in the Atmospheric Sciences program here at Princeton, with a Sun 3/280 file server also on the net. I have the Apple ethernet cards in the machines, and am running NCSA Telnet 2.3 (which has server FTP support). Basically, between the mac and the server I am getting no better than about 35K bytes/sec for binary transfers, no matter how I tweak the protocol parameters. On the same net, Sun 3/50's routinely do about 70K bytes/sec at the same time. On the TCP/IP scorecard posted here earlier, I have seen speeds of up to 100K bytes/sec posted for IBM PC ftp speeds. Are my experiences typical? Is there anything I can do to speed things up? Is the problem in software or in hardware?
amanda@intercon.uu.net (Amanda Walker) (07/24/89)
In article <9559@phoenix.Princeton.EDU>, rpbert@phoenix.Princeton.EDU (Raymond Pierrehumbert) writes: > Can somebody tell me what the bottleneck is on FTP transfer rates for > a MacII on ethernet? There are a number of factors. The biggest ones are (in no particular order): - SCSI Disk speed: You will get a speed improvement by using a good disk cache or transferring to or from a RAMdisk. - Ethernet board & driver: The faster the board (and .ENET driver) can move bits around, the better. The fastest I have seen in action personally are the Dove FastNet III and the Asante MacCon II/E. - TCP/IP implementation: The native NCSA TCP/IP kernel runs in what would be "user time" under UNIX. In other words, most of the protocol processing happens during the main event loop. MacTCP, on the other hand, does most of its processing at interrupt time, which means among other things that ACKs go out immediately, which keeps the other end's idea of the RTT low. Using MacTCP will give you a significant performance improvement "right out of the box." Hope this helps, -- Amanda Walker InterCon Systems Corporation -- amanda@intercon.uu.net | ...!uunet!intercon!amanda
bzs@ENCORE.COM (Barry Shein) (07/25/89)
>Can somebody tell me what the bottleneck is on FTP transfer rates for >a MacII on ethernet? I am running two MacII's on a subnet in the >Atmospheric Sciences program here at Princeton, with a Sun 3/280 >file server also on the net. I have the Apple ethernet cards in >the machines, and am running NCSA Telnet 2.3 (which has server >FTP support). Basically, between the mac and the server I am >getting no better than about 35K bytes/sec for binary transfers, >no matter how I tweak the protocol parameters. I believe if you run benchmarks writing the disk locally you'll find it peaks at around 50K bytes/sec. With the additional overhead of the network activity (those disks are all PIO, right?) you're probably doing well at those speeds. -Barry Shein Software Tool & Die, Purveyors to the Trade 1330 Beacon Street, Brookline, MA 02146, (617) 739-0202
amanda@intercon.uu.net (Amanda Walker) (08/01/89)
In article <8907250124.AA23044@multimax.encore.com>, bzs@ENCORE.COM (Barry Shein) writes: > With the additional overhead of the > network activity (those disks are all PIO, right?) you're probably > doing well at those speeds. Yup. I *really, really* wish Apple would let us do async I/O to the file system. I can ask for it, but the OS ignores me. Grumble, grumble. Sometimes you *want* to treat the disk as a slow device... -- Amanda Walker InterCon Systems Corporation -- amanda@intercon.uu.net | ...!uunet!intercon!amanda
desnoyer@apple.com (Peter Desnoyers) (08/01/89)
In article <8907250124.AA23044@multimax.encore.com> bzs@ENCORE.COM (Barry Shein) writes: > >Can somebody tell me what the bottleneck is on FTP transfer rates for > >a MacII on ethernet? I am running two MacII's on a subnet in the > >Atmospheric Sciences program here at Princeton, with a Sun 3/280 > >file server also on the net. I have the Apple ethernet cards in > >the machines, and am running NCSA Telnet 2.3 (which has server > >FTP support). Basically, between the mac and the server I am > >getting no better than about 35K bytes/sec for binary transfers, > >no matter how I tweak the protocol parameters. The numbers I have seen for memory-to-memory MacTCP and AppleTalk transaction protocol performance are very close. I would suspect that the actual hardware driver that receives the packet (and is common to both stacks) is the bottleneck. That bottleneck is about an order of magnitude faster than the FTP performance I see to a Sun (about 35kbyte/s, like Barry). > I believe if you run benchmarks writing the disk locally you'll find > it peaks at around 50K bytes/sec. With the additional overhead of the > network activity (those disks are all PIO, right?) you're probably > doing well at those speeds. My SC80 runs considerably faster than that. (I did a quick-and-dirty benchmark just now and got ~150kbyte/s to duplicate a file. One-way performance should be better.) My guess is that the bottleneck is in NCSA Telnet, or at least in its interface with MacTCP and the file system. Peter Desnoyers Apple ATG (408) 974-4469 Disclaimer - I make no claims about the performance or any other characteristics of Apple products. These figures are for comment only. Also, I don't mean to put down the fine efforts of the people who have produced NCSA Telnet.
amanda@intercon.uu.net (Amanda Walker) (08/02/89)
In article <3258@internal.Apple.COM>, desnoyer@apple.com (Peter Desnoyers) writes: > My guess is that the bottleneck is in > NCSA Telnet, or at least in its interface with MacTCP and the file system. As I've already noted, I'd say it's mostly in the latter. NCSA blocks disk I/O in 8K blocks, which is pretty reasonable, but it still blocks. The MacTCP interface routines, while not the absolutely most efficient (in particular, they use TCPRcv instead of TCPNoCopyRcv), aren not bad, and they certainly don't account for the 80% difference in performance between your memory-to-memory experiment and the observed FTP performance. Using the same MacTCP interface but not writing stuff to disk, I have no problem getting 100-150K bytes/second. Another thing that can make for a big performance hit is running under MultiFinder, since this reduces the percentage of time in which the FTP server can process data and write it to the disk. I'm glad that the file system will support asynchronous I/O in System 7.0; that'll help this sort of thing a lot. -- Amanda Walker InterCon Systems Corporation amanda@intercon.uu.net | ...!uunet!intercon!amanda -- "We may have come here on different ships, but we're in the same boat now." --Betsy Rose
pat@grebyn.com (Pat Bahn) (08/11/89)
In article <3258@internal.Apple.COM> desnoyer@apple.com (Peter Desnoyers) writes: >In article <8907250124.AA23044@multimax.encore.com> bzs@ENCORE.COM (Barry >Shein) writes: >> >Can somebody tell me what the bottleneck is on FTP transfer rates for >> >a MacII on ethernet? I am running two MacII's on a subnet in the > >The numbers I have seen for memory-to-memory MacTCP and AppleTalk >transaction protocol performance are very close. I would suspect that >the actual hardware driver that receives the packet (and is common >to both stacks) is the bottleneck. > >That bottleneck is about an order of magnitude faster than the FTP >performance I see to a Sun (about 35kbyte/s, like Barry). > Barry You are doing approximately 3.5 times faster on your FTP then I was doing on a benchmark we had using a mac2x and a mac2. We felt the bottleneck was the disks as using an excelan LANALYZER we observed peak performance in excess of 100kbs so the card and memory were capable of quite a good clip. The only parameters that might help may involve increasing the memory buffers FTP uses. We had notoriously slow disks on our system. Unless you want to buy some disks there is nothing you can do. AS an experiment get some software for a turbo disk (quasidisk, whatever) and see if that helps you out. I would not expect you to ever beat 100kbs though. If you do let me know I am very interested. >> I believe if you run benchmarks writing the disk locally you'll find >> it peaks at around 50K bytes/sec. With the additional overhead of the >> network activity (those disks are all PIO, right?) you're probably >> doing well at those speeds. >My SC80 runs considerably faster than that. (I did a quick-and-dirty >benchmark just now and got ~150kbyte/s to duplicate a file. One-way >performance should be better.) My guess is that the bottleneck is in >NCSA Telnet, or at least in its interface with MacTCP and the file system. > Peter Desnoyers > I have one caveat, the work I was doing involved AUX on both systems. I would expect MACOS stuff to run faster as the system is more mature and has a better design. But I would look at the disks as the problem. We felt the bus and memory were more then fast enough and that the CPU had the throughput. But we saw a large variance as we changed disks around. I did not have the documentation on how to tune FTP so I don't know what was possible, but I don't think the MACTCP or the file system is it. Pat B Stuff follows to defeat inews. type n to escape. A A A A A -- ============================================================================= Pat @ grebyn.com | If the human mind was simple enough to understand, 301-948-8142 | We'd be too simple to understand it. =============================================================================