As IT adds I/O and communication-intensive apps to servers, the strain on data center networking increases drastically.
There's more than one way to break a log jam of network traffic, from inexpensive and quick fixes to strategic capital investments and restructuring. New technology is set to boost networking capacity, and network fabric adoption has solved backbone performance bottlenecks with multi-lane Ethernet. In some cases, simple operational organization relieves congestion.
Storage with servers
Google colocates networked storage nodes with multiple servers, and then arranges the apps to use data from the nearby storage. This approach allows you to add inexpensive ports to the in-rack switches, or even have two switches, which allows dual Ethernet ports on the nodes. It also makes it easy to have four or more ports on the storage nodes, removing a network bottleneck to allow data in and out. Almost all of the data flow in the rack is through top-of-rack switching -- at low latency -- and backbone traffic is dramatically reduced.
Databases are different. The most efficient models today use a large pool of dynamic random access memory (DRAM) dual in-line memory modules to create an in-memory database. Ideally, the IT organization buys a tailored new server fleet -- some of which can hold as much as 6 TB of DRAM -- although older servers work too.
An adjunct to the in-memory architecture is to include fast solid state drive (SSD) storage in the servers. This can be used as staging buffers for the DRAM or as a networked storage resource. Both approaches reduce network load, but 10 GbE networks are probably too slow to keep up with systems less than one year old, even with two ports per server.
Virtualization is common in x86 server clusters, bringing its own network bottleneck problems. Boot storms are infamous for saturating networks; even in steady state operation, creating instances adds a load, as gigabytes of data shift from networked storage to the server.
In this situation, transfer allegiance from traditional virtualization to the container model. This means giving up the flexibility to create an instance with any OS, but that's usually not an issue.
The container approach, which reduces network traffic, requires every instance in a server to use the same (container-supporting) OS. DRAM space saved from a single OS and app stack allows double the instance count, and startup is quick. However, if the apps in the instances are network- or I/O-intensive, the same stresses can occur.
Technical fixes of the future
Inter-switch connection using 40 GbE (quad-lane) links is common, and we are looking for 100 GbE as an alternative to 10 GbE. This trend is getting a boost from efforts to deploy 25 GbE links, which will allow four-lane, and relatively inexpensive, 100 GbE links for storage devices and inter-switch connection.
With 25 GbE, data centers use existing cabling in the rack and between switches. Unfortunately, you cannot retrofit adapters, instead using a PCIe card or a new node. Even so, replacing top-of-rack switches to create a 10/100 GbE environment as soon as economically possible will substantially boost overall cluster performance.
This new technology is rapidly moving into the market, reflecting the needs of cloud service providers. The projects to deliver 25 GbE links are generally less than 12 months old, and IEEE ratification is set for a record schedule. Production network interface cards and switches are expected in the second half of 2015.
There is also a 50 GbE dual-lane option in the pipeline. The faster speed could fit larger servers running in-memory big data analytics, probably with at least two links per server. With the trend for these servers and high-performance computing to have massive core count CPUs or GPUs, expect data starvation to be an issue, as will the time to load terabytes of memory with data.
Software-based fixes also address bottlenecks. Software-defined networking can spread the load of tasks on backbone lines across servers.
With storage and architecture performance growing rapidly, networking is going to be at the forefront of innovation for the next decade, so evolution should be rapid.
About the author:
Jim O'Reilly is a consultant focused on storage and cloud computing. He was vice president of engineering at Germane Systems, where he created ruggedized servers and storage for the U.S. submarine fleet. He has also held senior management positions at SGI/Rackable and Verari; was CEO at startups Scalant and CDS; headed operations at PC Brand and Metalithic; and led major divisions of Memorex-Telex and NCR, where his team developed the first SCSI ASIC, now in the Smithsonian.