I don't fully trust Tom's Hardware to begin with. They've been known in the
past to skew results. I personally wouldn't use WinXp in anything
mission-critical to begin with, so the point is moot. They offer nothing
towards the Linux crowd.
The numbers they provide in that article make sense, except they're
about 20% lower than the best I can get from Linux. The ones on the
linked page are quite close to what I've seen in Linux if the default
TCP buffer parameters parameters (wmem = 4096 16384 131072, rmem =
4096 87380 174760) are used. The defaults buffer sizes - 16384 and
87380 are a bit too small for a single connection.
However, PCI is amazingly slow in comparison to everything else. With PCI
Express moving in, all these old standards will move out. PCI-X is nice and
all, but as far as expansion goes it's still kind of limited. All the other
PCI specs are just slow (32-bit, 64-bit, 33mhz, 66mhz, all slow). PCI Express
is now being used on some motherboards, you might want to check out the speed
there... I've heard/seen that the performance for those is amazing...
One of the nicest parts of PCI Express is that it enables a crossbar
switch between peripherals. Every device is on its own segment.
Performance is fast, predictable, and good.
However, if CSA is that important, grap a linux box and hack it apart...some
real raw numbers are always pretty, even if they aren't always real-world...
The nice thing about CSA is that it's been around long enough to be
cheap and widely available. It likely lets you put together a rather
powerful file server or iSCSI target device on the cheap - all disk and
network traffic move through only the *CH hubs without ever touching
PCI. PCI Express hasn't arrived yet, but once it does CSA will go away.
Alan
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html