This article is more than 1 year old
Intel's new Xeon: Easy to switch between dual memory modes? Uh, no
Resilience or throughput – set it then forget it
Analysis Intel's new Xeon E7 v2 server chips are capable of two different memory modes, but don't expect to switch between them willy-nilly or to have two VMs running on the same box using different modes.
As we noted in our deep dive into the innards of Chipzilla's new top-of-the-line x86 server processors, the Xeon E7 v2 series provides a dual-mode memory capability that's enabled by Intel's new JordanCreek memory extension buffer chips.
Before we get into that dual-mode capability, let's first take a look at the lay of the land. Each socket into which the Xeon E7 v2s are installed can support four JordanCreek chips; each JordanCreek can support two "next-generation scalable memory interconnect channels" – SMI Gen2, in less-of-a-mouthful parlance – that can each handle three DDR3 DIMMs. That's six DIMMs per SMI Gen2 channel, totaling a maximum of 24 DIMMs per socket, or 96 DIMMs in a four-socket configuration.
Still with us? Good. Now, the dual-mode nature of the SMI Gen2 channels allows for the memory interconnect to operate at either a 1:1 ratio with the DRAM clock in what Intel calls "Lockstep Mode", or a 2:1 ratio, in what they call "Performance Mode", "Independent Channel Mode", or "Two-to-One Mode" — take your pick.
Lockstep Mode supports 1600MHz DIMMS; Performance Mode (et al) supports 1333MHz DIMMs – but since the latter can transfer data at a 2:1 ratio, its effective memory bandwidth is 2667MHz. Lockstep Mode, as you might imagine, is to be preferred for workloads in which the highest degree of accuracy is required – think high-speed trading or financial transactions, for example – while Performance Mode is aimed workloads in which high throughput is of more value. Media streaming, anyone?
Now here's the rub: these modes are enabled at the firmware level after power-on, and once you've got your server up and running, there's no switching from mode to mode, let alone allowing, say, one virtual machine to use Lockstep Mode and another to use Performance Mode – it's "You pays yer money and you makes yer choice."
Lockstep Mode takes advantage of mirroring, an Intel rep told us after the Xeon E7 v2 series rollout event in San Francisco on Tuesday. "There's more reliability associated with it because you have more mirroring associated with what's happening in the memory system on the Lockstep Mode," he said. "So you're not going to get as much performance, but you're going to get more resilience."
I feel the need ... the need for speed
In Performance Mode, he told us, "You're going to have more addressability, more access to a bigger memory footprint in a non–Lockstep Mode. You're also going to have higher throughput."
The mode is established for the entire system at boot-up, and remains in effect throughout the server's uptime and for all processes running on that machine. This choice becomes important when you consider the increasing interest in in-memory computing, a core theme of the roll-out presentation by Intel SVP and GM of the data center group, Diane Bryant, who touted the large memory footprint of servers based on Xeon E7 v2 series chips.
Working on large data sets entirely resident in memory rather than having to swap data in and out of memory from SSDs or even – shudder – hard drives can result in phenomenal performance gains, she claimed, citing boosts in the 125X to 150X range.
Should you choose to set up your Xeon E7 v2–based server to perform in-memory computing, know that once you choose its memory mode, you're essentially stuck with it unless you take the server down and reconfigure in the firmware settings.
Naturally, we were interested in how long such a reconfiguration might take, since downtime is the bane of any system admin – or, more to the point, bean-counting CIO. But when we asked HP worldwide product manager Tushar Agrawal to give us a ballpark estimate – hours? days? – of how much downtime would be involved in such a switcheroo, he declined to provide an estimate.
"It's hard to put a number like that" on such as estimate, Agrawal told us. "Downtime is really dependent upon a lot of different factors, so it's hard for me to say what it would be."
Fair enough. But you, dear Reg reader, would be wise to carefully test and optimize your brand-spanking new Xeon E7 v2–based server before putting it into production. Once it's up and running in either Lockstep Mode or Performance mode, odds are you're not going to want to take it down.
That said, from Agrawal's point of view, your choice of memory mode may not be that big a deal. "I think for most of the workloads, they're probably not going to see a huge impact," he said.
So from what we were able to learn during this morning's event, if you plan to step into the brave new world of in-memory computing, you could lean a little towards resiliency or a little towards throughput.
Or you could just flip a coin. ®