On Saturday 31 July 2004 12:04, Alan Cox wrote:

On Sat, Jul 31, 2004 at 11:29:50AM -0400, Lamar Owen wrote:

...

and it has CONFIG_ATM unset, too. My question to the kernel maintainers here (Arjan, Alan, and Dave) is simply 'Why no ATM?' The ATM support is actively maintained, is production quality, etc.

That is good to hear. Can you tell us which cards and setups you've tested

I currently have a Dell PowerEdge 1600SC (2.4GHz Xeon, 1GB RAM, 66MHz 64-bit PCI slots) running the FC1 kernel with the shipped he drivers, linux-atm-2.4.1, and a Fore HE622 OC-12 card connected to a 3Com CoreBuilder 7000 OC-12 interface.

I know the kernel configurator says it's EXPERIMENTAL; I'm just not sure that's the case any more.

I have many more cards; I am able to test Fore PCA-200, Fore LE, IBM Interphase 5155, and a few others. For obvious reasons the HE622 support is the most important to me for server/firewall use.

Chas Williams is the HE maintainer for both the 2.4.x series kernels and the 2.6.x series kernels, and has updated the he driver pretty well during 2.5.x and 2.6.x development. While the userland tools have not had a release in a year, the tools are fairly stable, and CVS commits have occurred pretty recently. It is fairly mature technology, judging by the linux-atm list traffic.

But, since the ATM configuration was not included in the FC2 kernels, it makes it difficult to test there. I am assuming that there was a good reason for not inlcuding the ATM config, but Arjan or Dave will have to answer that question, as the changelog in the FC2 kernel doesn't mention a reason, nor does it mention that the CONFIG_ATM option is even unset. It doesn't even mention ATM at all....

...

I am looking at system-config-network too for ATM support; I know this is not a RedHat-supported item, but it is something I will need at some point, since I want native ATM networking available to some of my users.

I can't see it ever becoming Red Hat supported (unless there is strong RHEL customer demand). For Fedora however the big problem is knowing if it actually works.

ADSL PPPoA connections are ATM; having full-bore tools can help there too, unless the PPPoA stuff is separately supported. (ADSL is a B-ISDN ATM technology, after all, and high-bandwidth Internet connections at OC-3, OC-12, and higher rates are typically POS (Packet over SONET) or ATM (my local ISP quotes OC-3 connections, and prefers ATM CLIP delivery)).

So ATM features are more Enterprise than hobbyist, but having robust ATM networking for CLIP over PVC and SVC as well as LANE support could raise some interest in high-bandwidth applications.

In my case, the server that the HE card is in needs to associate with multiple ELANS for firewall and routing purposes; ATM NIC's are great for that sort of thing, since I can set up a CLIP VC and two or more LEC's on the one card; but as far as the kernel is concerned they are physically separate 'ethernet' interfaces.

The problem ATM has faced in the past is a combination of cost and complexity. Cost is going down, and enterprise-class switches routinely turn up on eBay at a fraction of their retail cost (hrmph, I've seen several Nexabit NX64000 switch/routers capable of OC-3072 internal speeds routing 6.4Tbps regularly cross eBay for ~$2,000, even Juniper and Cisco equipment at OC-12 and OC-48 speeds are relatively affordable). The complexity is a turn-off, but a good GUI to the linux-atm stack and good support of that stack could flatten the learning curve substantially. After all, without any help, I brought up a five-switch fully redundant ATM network with multiple ELANS and LECS/LES/BUS automatic failover in just a few months (and, while my title is IT Director, I am basically the entire IT department, and do everything from terminating fibers to writing software). It just wasn't that hard. And it impresses people to no end when I physically pull an OC-12 trunk fiber out of a switch port and the switch transparently reroutes over the other fibers....

On Saturday 31 July 2004 13:51, Charles R. Anderson wrote:

On Sat, Jul 31, 2004 at 12:46:53PM -0400, Lamar Owen wrote:

...

So ATM features are more Enterprise than hobbyist, but having robust ATM networking for CLIP over PVC and SVC as well as LANE support could raise some interest in high-bandwidth applications.

ATM is still too overly complex as an Enterprise networking technology. Certainly support should be there for those who need it, but I doubt it is worth the effort just for new high-bandwidth applications, unless the intent is to work with legacy ATM infrastructure.

But all telco infrastructure on the WAN side is SONET, and most is ATM.

Ethernet can do that with 802.1Q VLANs. I just don't see the advantage of ATM for new deployments in this case.

Real traffic isolation at the link level. Since ATM is connection-oriented, there are no broadcasts to sniff, no way for a random user to get any traffic they aren't supposed to see.

There are a few reasons I picked ATM over gigabit Ethernet. The first one was distance. My fiber runs average longer than 300 meters, which is the limit for 1000Base-SX on my fiber (which is already installed, and relatively old at this point, having been installed in the early to mid 90's. I have more than one run of fiber that I am using that is longer than 600 meters, but works fine with OC-12 over multimode (HP HFBR 5208 transcievers)). 1000Base-LX is too expensive.

I have $2,500 invested in my ATM LAN with OC-12 links. Server connections are available with 1000Base-SX and OC-12 ports. The switches feature full hot-swap redundant power supplies, hot-swap interface cards, and enterprise SNMP management.

Can 802.1Q give me full, automatic, link failover at the link level? With automatic load balancing? The PNNI implementation in the 3Com gear does both, transparently to the AAL5 above.

Emulating broadcast LANs on top of circuit-switched technology is inefficient, complex, and fragile. On the other hand, the support could be useful for those stuck with legacy ATM in order to create firewalls, IDS's, flow analysis tools, etc.

Yes, LANE is not the most elegant solution ever invented. But once the VCC is up, the throughout is very good, and competive to CLIP. As far as 'legacy' goes, ATM is at the core of new technology, not just old. ADSL, for instance, is ATM. To support some USB ADSL modems requires the ATM stack; in particular, PPPoA on the Alcatel Speedtouch is in kernel-2.6, but not built in the Fedora kernel by default.

Ugh. It seems silly to create all that complexity just to work around the inherent weaknesses of the LANE technology, with all the overhead to boot. I don't think you characterize the average IT person.

No, I don't, I guess. 3Com's Transcend Enterprise Manager did all the heavy lifting. I just had to think through what I wanted, and learn the terminology, and characterize the topology desired. PNNI on those switches isn't hard, once you wrap your mind around it, remembering that ATM is circuit-switched technology. And you don't have to do anything at the IP level; it's all handled transparently to the IP layer.

I have yet to find an affordable gigabit layer 3 switch that can do now what this ATM gear could do in 1998. Automatic load balancing and failover were and are dealbreakers for me, given the geographic issues on my campus. Now, these ATM switches were not really affordable in their day; but they are _now_, and on an educational budget they work wonderfully well.

...

And it impresses people to no end when I physically pull an OC-12 trunk fiber out of a switch port and the switch transparently reroutes over the other fibers....

Ethernet can do that too with 802.3ad Link Aggregation. I too went through my ATM phase and I'm much happier now that I have migrated to a completely Ethernet infrastructure.

I had a couple of Bay Networks switches that did the aggregation and trunking. It was not as smooth nor as fast. Since implementing the ATM LAN things are much smoother here. YMMV, of course. ATM is not the be all end all, but neither is Ethernet. Nor is TCP/IP for that matter.

However, ATM is still very popular, in particular for the home user who wants to use Linux and a USB ADSL modem that implements PPPoA (Alcatel Speedtouch, for instance). This is not uncommon, and can provide lower latency and higher throughput than PPPoE. I know a number of users on my ISP that run the Speedtouch on Windows; they can't switch to Linux unless the PPPoA support is there.

But the biggest reason I like and use ATM is the fact that there is no boundary between WAN and LAN, since ATM is mostly a WAN technology. And I am going to be using a Linux ATM firewall when I get a bigger pipe out of here. The usage to which I am going to put my facility also make ATM attractive; I want to be able to park a cage in a building and give users their own PVC's in that are invisible to other users here, even though they are carried by the same fiber and the same switches.

But our goal here is an OC-48 incoming for real-time telescope control; not the typical application, granted.

My new ATM-enabled kernel is up and running on FC2, now. Only took 90 minutes to rebuild the RPMs on a 2.4GHz Xeon (!!!???). Now to test my cloud's bandwidth.

On Wednesday 04 August 2004 16:37, Arjan van de Ven wrote:

On Wed, 2004-08-04 at 22:29, Lamar Owen wrote:

...

Arjan, I've not seen a comment from you on this. Can you please comment, even if the comment is for me to just go away with my ATM junk? :-)

ATM is of, ehm, mixed quality. Some parts are ok, some are undermaintained it seems and might pose a security risk. Note that the current erratum kernel has ATM at least partially turned on but I would appreciate feedback on which hardware works to be able to make a more informed decision about what to enable exactly... "it compiles" vs "it gives scary warnings" isn't the bes ;)

First, thank you for taking the time out of your busy schedule to reply. Second, the HE drivers seem to be well-maintained, with Chas Williams being active in patching as well as on the linux-atm list. The HE622 is the hardware I am using currently in a couple of Dell 2.4GHz Xeon servers (ServerWorks chipset). I have done some stress testing using ttcp amongst others, and the HE is giving high throughput numbers on both the FC1 kernel and the FC2 kernel (my custom compiled version, based on the last production erratum). I have stress-tested over several hours, and system load stays pretty low, and things aren't throwing warnings or 'scary' errors.

In the interest of helping the project, I will be testing other adapters shortly. Mike Westall of Clemson University can speak to the Interphase and Turboways drivers and general stack stability, as he has developed and maintained an ATM research network for several years (since 1997, I believe) and he is also available on the linux-atm list. He also is currently active.

When building my custom kernel I did see some fairly scary warnings from one of the drivers, but I forget ATM which one it was...Zeitnet?

The userland is a pill to build on FC2, unfortunately, but that isn't relevant in this context.