Most users tend to think of ATM in terms of 155 Mbps connections. For several reasons, ATM at 25.6 Mbps (ATM 25.6) is better suited for most desktop connections than is ATM at 155 Mbps (ATM 155). This paper explores these reasons.

Shared networks are running out of gas. Nobody knows when the network bandwidth will run out, especially with Ethernet. How many network managers know the average bandwidth utilization of their networks? If their Ethernet network segments are averaging over 40% utilization, then next Friday at 2:00 p.m. they might just stop completely!

There was a time not so long ago when everyone thought 640K of memory in a PC would never be used up. Now it would be hard to buy a new machine with so little memory. There was a time when shared use of 4 Mbps or 10 or even 16 seemed infinite. How could we ever use that much speed in an office of small PCs?

There is a lot of talk about finding the killer application that will bring everyone rushing to ATM the way VisiCalc did for the original Apple computer. Let’s face it, not everybody is rushing out to buy the video hardware necessary to run CU-SeeMe over the Intranet. While everyone has been looking, however, the killer apps are here. Small changes in usage patterns of existing tools and a few new tools with bandwidth needs only slightly bigger that what we have been using before have brought some real networks to their knees:

A Texas software firm developed a spiffy new screen saver mode for their product. This feature involved displaying some fancy moving graphics on the monitor. This was fine in a normal situation, but in their office they had a 100 or so people running a beta test version off of a single file server. The first day the new version was up and everyone went to lunch, the network collapsed.

• A California utility had an email network throughout the state that was running fine. One day, users discovered that they could attach graphic images to what had been short little text email messages. All of a sudden the network stopped because the required bandwidth jumped an order of magnitude.

• Financial institutions often bring up an application that broadcasts a stock ticker that any user can pick up without having to log on to the application. The broadcasts go everywhere, and any WAN line at 56Kbps or less will be passing little else besides the ticker traffic.

• Users put new applications into production or change to a GUI look for existing applications. These trivial seeming changes will use lots more bandwidth or otherwise change the traffic mix of the existing client/server paradigm. As these changes are brought on-line, what will happen to the network?

There are several aspects of ATM technology that make it highly attractive:

• Higher bandwidth -- ATM has plenty, and there is no end in sight. Other technologies also promise higher bandwidth, however, so this is not an exclusive for ATM, but ATM does supply it.

• Homogeneous, scalable bandwidth -- One of the most important points about ATM is that it allows networks that are homogenous from end-to-end. ATM is capable of scaling all the way from T1 (or even lower) to multi-Gigabit speeds without requiring any format conversions on the way. Such networks are dramatically easier to administer than are networks containing a mixture of technologies.
• End-to-end flow control -- ATM includes hardware level features for using flow control to accommodate congestion in the network. These mechanisms work best only when they are extended back to the source host. ATM host attachment therefore allows very large networks to work better than networks where the flow control is done several layers higher in the Transport layer. The WAN will almost certainly be ATM in the future if it is not already. It makes good sense to extend this technology back into the LAN.
• Quality of Service for voice, video, and data -- ATM works its magic by chopping up larger blocks of information into relatively small cells. These small cells have two advantages: they can be switched cheaply and quickly with hardware, and large data packets cannot block the smooth continuous flow required for voice and video. Putting all these types of traffic on one network lowers recurring WAN line costs and simplifies management.
• Shared line to the TELCO -- Frame Relay provides multiple data connections to the local TELCO over one line, resulting in significant savings. Again, ATM also allows voice and video to ride on the same lines as the data, resulting in even greater savings, and in the long run, better manageability.
• Internetworking -- legacy bridge products have limits that make it difficult to deploy large networks. LAN Emulation and new broadcast management techniques can overcome many of these limits. Routing is relatively slow and is hard to configure and manage. ATM, more like bridging, is easy to install and manage. This helps the network manager cope with reduced staff and increasing workloads. Products on the horizon will merge Routing with Switching, using Routing for short interchanges and switching for longer connections.

Consider the following typical demo setup. Two Pentium PCs with ATM NIC's are set up with a VCC through an ATM switch at 155 Mbps. Both are running Windows NT software. One is acting as a server, and the other is reading an AVI file (a Motion JPEG video format) and displaying it on the screen. The size of the video is only about 4" square. The bandwidth used in this setup is about 1 Mbps from server to client and is almost 0 in the other direction. However, the client is running at about 25% of its maximum CPU utilization.

The point is that for PC class machines today, ATM 155 Mbps is a huge overkill. (So is switched 100BaseTX.) Using it on a PC would be rather like trying to drink out of a fire hose. Incidentally, the server machine is only running 1-2 % CPU utilization, indicating that a file server probably could reasonably use a 155 Mbps ATM connection.

The following table shows Maximum burst speeds for various PC buses. Note that when data comes in to a PC from the network, it must cross the bus at least twice, once going from the NIC to RAM and then from RAM to wherever it will be used -- most likely the CPU, the hard drive or the video.  

The PCI bus has two variables, bus size and speed. The original version is shown as PCI. NIC's are now available with 64 bit width (shown as PCI-2) and with 64 MHz speed (PCI-3). The Sun Systems SBus is shown for comparison.

In addition, bus cycles are also taken for the running program and any other data it is using. This is why these are "burst" speeds. The PC can not sustain this speed indefinitely on a single task. In addition, newer operating systems are multitasking, so presumably there are other things going on, background printing, email updates coming in, etc. Today’s buses are not well suited to sustained speeds like 100 or 155 Mbps.

The next table is taken from a Bay Networks presentation. It shows bandwidth requirements for typical applications, especially some touted as driving the need for ATM. As you can see, there is little that will require anything like 25.6 Mbps, much less switched 100 Mbps Ethernet or 155 Mbps ATM.

Think about 25 Mbps of bandwidth. At that speed you can fill today’s typical 1 GByte hard drive in just over 5 minutes. What will the network do the rest of the day? Think about an engineer downloading a 100 M Byte file to work with. At 25.6 Mbps that transfer would take about 30 seconds. At 155 Mbps it would take about 6 seconds. Is the time difference worth the price difference? A user watching a high quality video only needs 15 Mbps. What else could he also be doing at the same time?

Prices for Competing Technologies -- Speed versus Cost

The next figure shows some approximate figures of price per connection versus performance in Mbps for the most likely competing technologies that bring more bandwidth to the desktop. In the figure, 100S means shared and 100X means switched. The pricing shown for 10 Mbps shared Ethernet is for inexpensive clone cards and hubs. When buying low end products, this is may be realistic. Users contemplating ATM probably would not buy this cheap a product. This low figure is used just to avoid quibbling about it. Shared Ethernet is quite cheap and it works fairly well. The pricing for the other Ethernet technologies is for name brand cards and switch, not a clone, taken from a retail price list. The 10 Mbps card price is probably low for this application.

A typical price is shown for a PC and an ATM 25.6 NIC and a 25.6 Mbps switch port, showing what is commonly called the "per seat" price. Note that the switch port costs about the same as the NIC. Together they cost about 5 to 20% as much as the PC. Manufacturing costs for these boards is currently under $50. It should come down further as the learning curve progresses.

A typical per seat price is also shown for a PC with an ATM 155 connection. Again, the switch port costs about the same as the NIC. Together they cost slightly less than an average PC. This number will also come down, but this kind of card needs either a fast microprocessor or complex ASIC's and a lot of RAM, or both on the NIC, thus the cost will always be high compared with the 25.6 NIC.

The dark areas at the end of the shared blocks in the figure indicate that under some situations users might get this bandwidth. For example, a CAD group on 100BaseT might routinely get nearly 100 Mbps for most transfers because the use is intermittent transfer of large files to and from the CAD servers, so the users do not tend to interfere with each other. For multimedia use, however, there will be multiple streams of data running at the same time for long periods. The bandwidth per user in this situation is less, perhaps much less, than that shown here because of the statistical nature of the CSMA/CD protocol.

As an aid to putting these prices in perspective the figure also shows a monthly cost for the user of the PC. This price is probably low. What if this system is for an Engineer? Of course, if the installation includes PCs for 1,000 users, or 10,000, it tends to add up, but so do the salaries of the users at the keyboards.

The SONET framing format used for ATM OC/3c is highly structured. A significant portion is taken up by overhead. The next figure comes from the ATM Forum UNI 3.0 book and is more schematic than physical, but the significance is that the 155 Mbps clock rate does not result in 155 Mbps of "goodput". The actual amount of data delivered is less than 140 Mbps. This figure does not include the roughly 10% overhead in ATM cell headers. That overhead is the same for all ATM, so it is ignored here.

In addition, there may be a time lag between when a cell arrives at the transmitter to be sent and the time when it is actually sent. This is because an entire SONET frame has to be sent as one piece. If the transmission of one frame has begun just as the cell arrives, the cell may have to wait for the entire frame to be sent and the next frame to be put together. This can result in extra jitter, Cell Delay Variation (CDV), in the arrival time of Constant Bit Rate (CBR) traffic. This jitter is disruptive to video and voice if enough of it accumulates over several switch hops.

ATM 25.6 Mbps has almost no framing beyond the cells. The signal is clocked by a chip originally used for Token Ring that runs at 32 Mbps. Token Ring used Manchester encoding with two clock cycles per bit, resulting in a 16 Mbps data rate.

The 25.6 signal is transmitted with 4b/5b encoding (4 data bits are converted to 5 bits on the line) for redundancy checking. This results in 4/5s of thirty-two, or 25.6 Mbps. Some combinations of the 5 bits are illegal and would never appear in valid data. Some of these illegal combinations are given special names, specifically J, K and T. The J and K values are used in pairs to indicate an idle state. As soon as the transmitter has a cell ready to send it will output a pair of T values, indicating a cell is about to start. It is immediately followed with a cell or a stream of as many cells as are ready to send. There is thus an extremely low (1-byte) delay in starting to send cells. This low delay means that less CDV is introduced into the transmission -- the better for voice and video.

As we have seen, ATM 25.6 is a much better speed match for the PC than ATM 155. Even 25.6 Mbps may be overkill for now, but that will be a usable speed for some years to come. There will probably always be places in the network that will not need more speed than this. We also saw that ATM 25.6 is a better cost match for a PC platform than is ATM 155 -- the network connection probably should not cost as much as the machine.

ATM 25.6 will run on Category 3 wire (Cat 4 or 5 as well, of course) with normal distances that would be installed today. ATM 155 will run on Cat 5 wire, but the distances are more limited. Users should not face wiring upgrades just to go to ATM. The sales term for this is "investment protection."

ATM 25.6 uses a timing chip that is an off-the-shelf Token Ring chip. This part is therefore proven, reliable, and relatively cheap. The rest of the 25.6 NIC can easily be made in a single chip. This is why the NIC can be so inexpensive. "Single chip" is a slight exaggeration, because most NIC's use the cheap Token Ring transceiver part, but the rest of the job (not counting ABR) can be done in a single, fairly simple ASIC. This one chip means lower real estate, heat and power requirements. Since the bandwidth is lower, there is less need for RAM on the NIC. All of this translates to higher port density and lower costs, both in the NIC and in the switch. It also means that the industry will eventually see a PC Card (formerly PCMCIA) NIC for ATM 25.6 but likely not for ATM 155.

IBM protocols (NetBIOS and SNA) are non-routable protocols. They must be bridged because some packets do not contain a network address. The conversion to ATM will be easy for users of these protocols, since their networking architecture already is designed for a flat network such as that used by ATM. An ATM network is flat because ATM switching works primarily at the data link layer, just as bridges do.

As was mentioned, the transceiver chip used in ATM 25.6 is the same as IBM customers are currently running on their Token Ring networks. Thus IBM knows that there should be no physical layer problems with plugging in ATM 25.6.

Unlike the Ethernet world, where Fast Ethernet and Gigabit Ethernet promises future bandwidth upgrades, there has been no clear Token Ring upgrade path. Per user switched Token Ring works, but Token Ring NIC's probably will not get much cheaper and neither will switch ports. There is now a 100 Mbps Token Ring consortium that promises 100 Mbps Token Ring in 1998, but that will be late for some users, and the complexity of Token Ring means that those NIC's will never be as cheap as NIC's for ATM 25.6 will.

IBM invented ATM 25.6 and steered it through the ATM Forum. Since IBM made this big push, 100BaseT has been very successful and will be a cheap solution to the desktop, so maybe IBM has backed off of ATM 25.6 somewhat. But, Ethernet is not Token Ring so a transition to that topology would not be painless. ATM 802.5 format LAN Emulation (LANE) can be an easy upgrade for Token Ring users.