DragonFly users List (threaded) for 2005-03
Re: Differences between AMD and Intel CPUs [was: Re: Dragonfly andHyperthreading....]
In a message dated 3/10/2005 9:52:12 AM Eastern Standard Time, "Thomas Edward Spanjaard" <t.e.spanjaard@xxxxxxxxxxxxxxxxxx> writes:
><EM1897@xxxxxxx> wrote in message news:02FB1A9B.3D7CFFE9.000424FF@xxxxxxxxxx
>> I'd much rather hear why you don't think that the bus
>> speed of the network adapter is a worthy criteria for
>> hardware selection than your opinion on my conduct.
>> Your reasoning is what gives you credibility; not your
>> name. It seems implausible that you don't understand
>> what a bottleneck a standard PCI bus is.
>Just as a sidenote, I've been able to get 12.3MiBps each out of two standard
>3c905ctx-m over a single PCI bus, with concurrent disk access (the
>controller where network data was flowing to was also on the same bus, old
>peecee). It just pays to investigate the options the hardware has for you,
>and busmastering PCI with DMA certainly ups the average attained speed on
>the 'slow' PCI bus. Of course, for Gigabit you are limited even by the
>theoretical maximum transfer rate, but there's PCI-X out there, as you
>mentioned, to alleviate that. When it comes to PCI-X, *all* boards are
>relatively expensive, not just the AMD-powered ones. For instance, take a
>look at the Tyan catalogue, it gives you a fairly good impression of server
>mainboard price differences accross the three (four including the Athlon MP)
>x86 server platforms.
>Knowing that, it sounds ridiculous to claim that you can get a 3.2GHz P4
>with a motherboard with PCI-X for under 300$.
Well sorry, but you are just completely wrong about
everything. I've already cited the P8SCi as an example,
Im sure you know how to use google to get prices.
Your "test" is also completely wrong, and your analysis
of the bus wrong as well. For example, passing 80Mb/s
through a box with a 33Mhz/32bit nic might use 50% of
the cpu, and you pop a pci-x card of the same type (say
the em driver) and it will drop to 1/3rd of the usage.
Saying that you can pass some amount of data is not a
useful analysis, because if you use 100% of the cpu
instead of 30% to achieve the throughput its worthy
of note. There are 1000s of I/O operations that are not happening in the background, and as the bus saturation increases,contention increases and those operations take longerand longer. Its not rocket science, its very easy to
look at the interrupt load with different kinds of
cards. Anyone that thinks the way you do just isn't
paying attention. But, its very typical of what I've
been saying. That is, that a lot of people who should
know better don't seem to understand the basics of
what the real bottlenecks are. It seems almost comical
that teams of engineers are rewriting an OS trying to
squeeze some cycles out of a CPU without any understanding
of the main bottleneck to networking performance.