Advanced Computing in the Age of AI | Tuesday, April 16, 2024

Juniper Wraps Smarter Fabrics Around Faster Switches 

If you have a need for networking speed, Juniper Networks is rolling out a new line of Ethernet switches in its QFX line that offer more bandwidth, more ports, and lower latencies than existing models and that go up against the Nexus 6000 switches from Cisco Systems.

The company is also offering a new kind of virtual fabric that will make it easier to manage large numbers of switches as a single unit, and talking about an even higher-level architecture it has put together called MetaFabric, and in some ways, the new fabric is more important than the hardware that makes use of it.

Jonathan Davidson, general manager of Juniper's Campus and Data Center Business Unit, tells EnterpriseTech that this ability to manage switches through Virtual Fabric (for relatively small collections of switches) or QFabric (a higher-end management fabric that scales a lot further) is key for extreme-scale datacenters.

Most large enterprises think in terms of pods for their collections of servers and the switch ports that link them together and to the outside world. Juniper has tens of thousands of customers using its switches, says Davidson, and at the largest enterprises, the typical pod size is on the order of 5,000 to 6,000 ports, and then they have multiple pods. At hyperscale datacenters, where they are often using sophisticated Layer 3 switching over Clos topologies, pod sizes can scale up to 20,000 servers in a pod with one or more ports per server. Juniper's QFabric software scales up to 6,000 10 Gb/sec ports in a single fabric, so you need multiple QFabric domains to scale like that.

Not every enterprise needs pod sizes that large, so Juniper created a lower-end offering called Virtual Fabric, which scaled up from four to ten racks of servers and up to 480 10 Gb/sec Ethernet ports in a single management domain. Virtual Fabric was simpler to implement and cheaper than QFabric and has been popular with many Juniper shops. But it didn't scale far enough.

So at the behest of customers, Juniper is offering a third new option, which merges some of the features of both of its existing fabric, which is called the Virtual Chassis Fabric. This will scale from ten to sixteen racks of servers and up to 768 10 Gb/sec Ethernet ports in a single pod, with promises of greater scalability beyond that in the coming quarter. Using the new QFX5100 switches (more on them below), the Virtual Chassis Fabric has as many as four QFX5100s acting as a spine and up to 18 acting as leaf nodes to create this fabric. You can weave older EX4300, QFX3500, and QFX3600 switches into the Virtual Chassis Fabric as well.

All three fabrics mentioned above will continue to be supported by Juniper, but it isn't hard to envision a day when Juniper has one product with one name that can scale up and down as needed.

Get Meta With Fabrics

The MetaFabric architecture that Juniper is launching is as much an attitude about how hardware and software will be created at Juniper as it is about any particular product. The idea is to provide open interfaces in Juniper software and to hook into open interfaces to third party products to create a higher-level fabric that allows customers options as they build out software-defined networks.

juniper-metafabric-block-diagram

So, for instance, under with MetaFabric, Juniper is supporting VMware's NSX controller, the OpenDaylight controller, and its own Contrail controller to manage the control planes in switches. In the same spirit of openness, Juniper's Contrail controller is able to interface with VMware's ESXi hypervisor now. The MetaFabric architecture also allows for Junipers switches to be linked into the OpenStack and CloudStack cloud orchestration tools. And like most such high-level architectures, MetaFabric 1.0 is coming with a reference implementation that puts Microsoft applications on IBM servers equipped with VMware hypervisors, backed by EMC storage and lashed by Juniper switches, routers, and security appliances.

Because we live in a big data world, MetaFabric has hooks to plug in analytics at all levels of the stack to monitor and control the network. One such tool is called Junos Space Network Director, which was initially created for managing campus networks and is now being applied to datacenter switches. It has a real-time performance analyzer for ports, VMs, and users and can create heatmaps of what is going on across devices. The VM analyzer actually keeps track of VMs as the live migrate around clusters of servers so you can do activity tracking at the VM and server level, and the fabric analyzer can monitor the health of Virtual Fabric, Virtual Chassis Fabric, and QFabric setups. Because the majority of enterprise customers are using VMware ESXi to virtualize their servers, Network Director plugs into VMware's vCenter console. Network Director costs $50 per device under management.

Going After Nexus 6000

That leaves the hardware. The new QFX5100 switches from Juniper are based on Broadcom's Trident-II+ ASIC, and Davidson says that these machines have double the density of 10 Gb/sec ports as their QFX3500 predecessors, have eight times the multicast routes, and, a port-to-port hop of 550 nanoseconds, have 50 percent lower latency.

The QFX5100 switch has a dual-core Intel processor running at 1.5 GHz plus 8 GB of memory and 32 GB of solid state storage to run applications alongside the Junos network operating system. The switch has two ASICs, each running a copy of Junos, and that means network admins can patch Junos on one side without taking down the other and losing a switch; once a patch is known to be working, they can route traffic through the updated half and patch the other half. This is called Topology-Independent In-Service Software Upgrade, or TISSU, for short. The switches support Virtual Extensible LAN (VXLAN), Network Virtualization using Generic Routing Encapsulation (NVGRE), and Open vSwitch Database (OVSDB) protocols, which are used to virtualize Layer 2 nets and extend them beyond the limits of VLANs.

The QFX5100-24S switch

The QFX5100-24Q switch

There are three models in the QFX-5100 family.

The QFX5100-48S is a 1U access switch that has 48 ports (SFP+) running at 10 Gb/sec; it has 1.44 Tb/sec of aggregate bandwidth and can process 1.08 billion packets per second. It has a suggested retail price of $30,000.

The QFX5100-24Q is also a 1U machine, and it can be used as an access or aggregation switch. This machine has its ASICs are revved up to 2.56 Tb/sec of switching bandwidth and has handle 1.44 billion packets per second. It comes with 24 QSFP+ ports running at 40 Gb/sec, and you can add up to two four-port expansion modules to push it up to 32 ports running at wire speed. This switch has a suggested price of $40,000.

The QFX5100-96S aggregation switch

The QFX5100-96S aggregation switch

The last machine, the QFX5100-96S, comes in a 2U chassis and is a high-density aggregation switch for 10 Gb/sec networks. It has 96 SFP/SFP+ ports plus another eight QSFP+ ports. It has the same ASIC configuration as the QFX5100-24Q above. Pricing for this switch was not revealed.

EnterpriseAI