Date: Mon, 05 Nov 2001 17:22:31 -0500 From: Bill Moran <wmoran@potentialtech.com> To: tlambert2@mindspring.com Cc: advocacy@freebsd.org Subject: Re: Article in pcmag Message-ID: <3BE71127.8030607@potentialtech.com> References: <3BE6BC06.356C4CE6@potentialtech.com> <3BE6F0BD.9F4173C4@mindspring.com>
next in thread | previous in thread | raw e-mail | index | archive | help
You don't have to email pcmag, Terry, but advocacy is advocacy, and (as some have stated) any PR is good PR. Terry Lambert wrote: >They didn't say which clients they were using; because of past >discussions, I'm going to bet that these were older clients, >and not NT Workstation/2000/XP clients, which perform significantly >worse under Linux than Windows, because of the data conversions >required. > I guess that's why there are lies, damn lies, and benchmarks ... You _are_ right, they absolutely should have described the clients used, not to do so taints the results by rendering them non-reproducable. They also leave out important details such as Linux & Samba version and W2K sp level. > >There is also some obvious Linux bias in the cut off on number >of clients; extrapolation of the curves shows at a small number >of addiioal clients, there are some additional heels in the >curves, and at least three of them have the Linux numbers so >close to converging on the Windows numbers, that the lines would >be bound to cross. > I'm not seeing the same thing in the graphs that you are. The response time graph shows nothing that you describe above, and I would argue that, while the graphs for transfer speed show a performance drop after a certain number of clients, your assumption that the lines will cross is presumptious. > It's also pretty clear that the limiting >factor was memory utilization of the per client fixed cost, and >that as soon as that was converted from cache to client state, >the Linux box augered in -- I epxect a FreeBSD box to do the >same, and the only fix would be to reduce per connection overhead >in the UNIX in question, and in the SAMBA server code itself. It >is pretty clear that NT has a bigger initial footprint, but the >per client footprint is smaller, and it's more efficient at the >data miss handling, where it has to load from disk, rather than >serve out of cache. > Except for the memory limiting issue, I don't see that at all. W2K with 512M RAM drops off at exactly the same # of clients as W2K with every other amount of RAM, whereas Linux does not drop off at all with 512M RAM. This (to me) says that the Linux/Samba combo is better at managing large amounts of RAM than W2K. The other RAM amounts show an equal "heel" (as you call it) for both Samba and W2K, the 128M downturns at 8 clients, while 256M downturns at 16 clients. How this means that W2K is better at data miss handling is beyond me. > >Perhaps the most telling point on where the heels were is the >amount of RAM in each box, relative to the heel-over. I don't >see a CPU utilization plot vs. number of clients, so the graph >seems to me to incorrectly correlate CPU speed, while glossing >over relative RAM size, which appears to be the determining >factor. From my own similar measurements in the past, I expect >that the CPUs are vastly overpowered for the NT, and might be >much closer to being consumed on the Linux side of things. > Well, I agree that RAM is the limiting factor, but I've never seen an NT machine with more effecient CPU usage than a Linux box, and certainly not one that had better CPU usage than FreeBSD. W2K I can't speak of as I haven't done enough experimenting with it to pretend to be expert. > >The bottom line is that I would not be happy to see FreeBSD >show similar curves to Linux, even if it beat Windows on the >limited number of clients in the test. > Well, the solution to that is to aid the coding effort to produce a better FreeBSD. And the main reason I put this on the list is that I _don't_ believe that FreeBSD would display the same curves. Especially with the new dirpref and dirhash code, I think FreeBSD would show much better curves. Obviously, you're not obligated to email pcmag, I just put this on the list in case anyone else might be interested. -- Bill Moran Potential Technology http://www.potentialtech.com To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-advocacy" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3BE71127.8030607>