RE: Clock Tolerance and WAN PHY
It would make our jobs easier (though less profitable) if we could plan for
next year's network development based on what computer systems do today.
Technology continues to evolve and there are some developments in progress
that will impact system bandwidth needs.
Today most large systems access their storage through Fibre Channel links
and access the rest of the world through Ethernet links. The Ethernet NICs
leave all or most of the TCP/IP handling to software which has to touch
every packet. Therefore, pushing a Gigabit data rate takes a lot of CPU
cycles. The Fibre Channel links do a lot of the packet processing in
hardware. In an unerrored transfer case, the software above a Fibre Channel
card gets messages built from many packets. They get up to Gigabit data rate
with much less system load.
People are developing iSCSI and iFCP protocols to support storage traffic
over Ethernet. In conjunction with that, there is also development of
hardware that will provide hardware functionality for these protocols
similar to that achieved by Fibre Channel. An Ethernet adapter may then be
used to support Ethernet traffic plus storage traffic with system
efficiency about that of FC.
This development can increase the data rate needs/capability of servers up
to the point where they need 10 Gbit.
From: Roy Bynum [mailto:rabynum@xxxxxxxxxxxxxx]
Sent: Monday, January 29, 2001 5:06 AM
To: rtaborek@xxxxxxxxxxxxx; HSSG
I am not sure if anyone has done in technical or economic study of how many
servers within the next two years will be able push close to 10Gb of data
into a LAN PHY. From what I have been told, computer systems today can not
even push enough to fill a GbE link, much less a 10GbE link. Without a
computer/server market, the 10GbE LAN PHY will be relegated to bandwidth
aggregation within privately owned campus fiber plant or a large data
room. The 10GbE WAN PHY, is easily used extended LANs over leased fiber;
MANs over privately owned or leased fiber and wavelengths; WANs over DWDM
wave lengths or leased fiber.
Given the amount of GbE that is being used by "legacy-free carriers" in the
U.S.A. and Europe, and the expansion of the Internet, the need for a
"service provider" type of Ethernet is becoming a major market
issue. Combined with the other uses of the 10GE WAN PHY, I believe that a
very high percentage (75%?) of 10GbE will be implemented using the WAN
PHY. The WAN PHY will be used because it has the operational management
functionality required. If P802.3ae does not go ahead and put the
operational management functionality that is being recommended, all of the
vendors will be implementing "proprietary" versions that will have it. If
P802.3ae does not do it, then the need to do it in proprietary
implementations will give 802.3 a "major black eye".
I tend to agree with Gary Nicholl in the need for a + 20 PPM
clock. Perhaps someone could give a relative cost to implement a + 20 PPM
clock instead of a + 100 PPM clock.
<Previous emails in the string deleted>