Date: Fri, 23 Apr 1999 20:06:57 -0700 From: Kent Stewart <kstewart@3-cities.com> To: "Joseph T. Lee" <nugundam@la.best.com> Cc: Michael Slater <mikey@iexpress.net.au>, "'freebsd-questions@freebsd.org'" <freebsd-questions@FreeBSD.ORG> Subject: Re: Fact or Fiction (Unix vs NT) Message-ID: <37213551.20FF811C@3-cities.com> References: <21EF26FF9AD8D01180E9BA3BC10000000EA13A@george1.iexpress.net.au> <3715886E.E6888C7D@3-cities.com> <19990423180711.A14514@la.best.com>
next in thread | previous in thread | raw e-mail | index | archive | help
"Joseph T. Lee" wrote: > > On Wed, Apr 14, 1999 at 11:34:22PM -0700, Kent Stewart wrote: > > I have a 100baseT network in my home. There are 4 Windows based > > systems (3NT and Win98). My hub shows >30% when I transfer files from > > one of my MS machine's to another MS machine using drag and drop. The > > %1 led lights up with FreeBSD. I see 200-400KB/sec (just like a few > > other people have been seeing on the list) when I FTP to my FreeBSD > > version 3.1 system. I tried ftp'ing files to my NT server. The only > > activity when I started the file transfers was an occasional > > heartbeat. What I recorded was 630-650KB for two files to NT. The same > > two files going to FreeBSD averaged 340-353KB. I have a few 16MB > > tarballs that I will try later. > > Well, you might have a misconfiguration somewhere. Here's my story of > how I got mine configured to maximum performance. > > I was working with my 100baseTX network last night and pulled 2.02MB using > FreeBSD -> FreeBSD (both 3.1-stable) and > 1.8MB using Win98 -> FreeBSD 3.1-stable, > by final configuration over leechftp in win98 and ncftp in FreeBSD. > > Hardware setup: > (1) server is a p133 running FreeBSD 3.1-stable with a > Netgear DEC-chipset 10/100 NIC. > (2) user machine is a celeron 450a running Win98 and FreeBSD 3.1-stable > using a 3com '905B NIC, on a ASUS P2B-S. > > It was pulling 88Kps before because I was forcing both NICs to go > full-duplex through a Netlux 10/100 hub which obviously wasn't handling > full-duplex itself. Switching to half-duplex got me the 1.8MB rate > from Win98 to fbsd transfering a 35MB file, over 3 tries. > > Then I rebooted the user machine to fbsd to check bandwidth there with > the xl0 driver. It was only pulling 300+KB with the same file transfer. > This was strange, and I wasn't aware of the xl0 driver having any problems. > > So while fiddling more, I kicked off a X11 build in the background. > Then, I noticed the ftp rate went up to 1.8MBps. > > I suspected that it was an IRQ matter, since the 3com card was plugged > into slot 4 which shares IRQ with SCSI and USB. It seemed that the > SCSI IRQs (during the X11 build) was allowing for faster transfers, so > I moved the 3com card to slot 1, which shares IRQs with the AGP card, > in order to feed the network card a regular stream of IRQs. > > I also went and idprio 0 the running rc5des program on both the server > and the user machine, although it actually had little effect on actual > transfer rates. > > The card move let transfer rates go up to a peak of 2.02MBps with low of > 1.8MBps. I'm not saying that the xl0 driver is sensitive to IRQs in a > certain way, but I'm saying that hardware configuration can affect > performance. > > I haven't tested the 3com card in the 2 unshared slots to really test > if unshared IRQs will affect it (a PCI128 and a Live! card sits in those > 2 slots). > My network card on the FreeBSD system is a de0 Dec clone. The NIC's in the Windows machines are all 3COM's. Four out of 5 are BusMaster 10/100 XL's and the fifth is a BusMaster Elink III 10/100. The FreeBSD system is a 166 MHz Pentium with 96MB of memory. The NT Server is a 2x133Mhz Pentium with 128MB of memory. I haven't rerun my tests where I was using the Elink III to talk to the de0 and the XL. The monitor is either the console for FreeBSD or the monitor for the Windows NT workstation. The console messages are more important right now. Your email caught me in the middle of some timed tests. A couple of people talked me into taking the step to a custom kernel, which led to a make world later. Since my results are significantly better, I thought I would let you know what I have found so far. I am also seeing a significant rate changes like you did but for a different reason. I finally figured out what was going on when I did a kernel make. My make times were significantly slower than a friend with about a 20% faster system. I started investigating to see what could be wrong with my hardware and found that the cpu cache had been turned off when I flashed the bios. The bios had problems with the larger drives and the latest version didn't. All of the bios parameters that I looked at hadn't changed but I hadn't looked at all of the parameters after the flash upgrade. The cache options were way down the screen. I recently did my first system upgrade and cvsup'ed to what I thought was 3.1 current. It was a kind of a shock because what I ended up with was 4.0 current. Before I move back to 3.1-current, I thought that I would run some simple tests to see what I would see right now. The generic 4.0-current kernel and system is running. I transfered a 26MB tarball file to my FreeBSD system from a P-II 400 running Win NT (XL > de0). The tarball is already compressed and we won't see any games played there. The transfer rate varied between 2.65-2.68 MB/s. This is much better than 400KB/s. The same 26MB file had a 1.24MB/s rate on a transfer to my NT server (XL > XL). There may be some parameter's that can be optimized on that system. In a short while, I will have a 200Mhz Pentium with 128MB running along side of the 166. Then I will be able to transfer files between identical OS'es to eliminate any bias there. My timed compiles on FreeBSD have dropped from 4961s to 406s. The FORTRAN build of MCNP is now 450s. It turns out that my hardware specific 3.1-release kernel was 15% slower than the 4.0-current generic kernel. This leaves the FreeBSD make running a little better than 2x slower. The scripts that came with the program don't provide a clean and the sample problems won't run if the output files already exist. I am working on that. All I can say is that FreeBSD is performing much better at this point. -- Kent Stewart Richland, WA mailto:kstewart@3-cities.com http://www.3-cities.com/~kstewart/index.html Hunting Archibald Stewart, b 1802 in Ballymena, Antrim Co., NIR http://www.3-cities.com/~kstewart/genealogy/archibald_stewart.html To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-questions" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?37213551.20FF811C>