Recently, I was part of a Composable Infrastructure briefing hosted by the enterprise side of HP and moderated by Tim Crawford. The event focused on a dynamically (re)configurable hardware. This type of innovation points to new generation of hardware with exciting capabilities; however, my takeaway is that we need more focus on the system level challenges of operations when designing units of infrastructure. I’d like to see more focus on “composable ops” than “composable hardware.”
What is “composable infrastructure?”
Basically, these new servers use a flexible interconnect bus between CPU/RAM/Disk/Network components that allows the operators to connect chassis’ internal commodities together on-the-fly. Conceptually, this allows operators to build fit-to-purpose servers via a chassis API. For example, an operator could build a 4-CPU high RAM system for VMs on even days and then reconfigure it as a single socket four drive big data node on the odd ones. That means we have to block out three extra CPUs and drives for this system even though it only needs them 50% of the time.
I’ve seen designs like this before. They are very cool to review but not practical in scale operations.
My first concern is the inventory packing problem. You cannot make physical resources truly dynamic because they have no option for oversubscription. Useful composable infrastructure will likely have both a lot of idle capacity to service requests and isolated idle resource fragments from building ad hoc servers. It’s a classic inventory design trade-off between capacity and density.
The pro-composable argument is that data centers already have a lot of idle capacity. While composable designs could allow tighter packing of resources; this misses the high cost of complexity in operations.
My second concern is complexity. What do scale operators want? Looking at Facebook’s open compute models, they operators want efficient, cheap and interchangeable parts. Scale operators are very clear: heterogeneity is expensive inside of a single infrastructure. It’s impossible to create a complete homogeneous data center so operators must be able to cope with variation. As a rule, finding ways to limit variation increases predictability and scale efficiency.
What about small and mid-tier operators? If they need purpose specific hardware then they will buy it. For general purpose tasks, virtual machines already offer composable flexibility and are much more portable. The primary lesson is that operational costs, both hard and soft, are much more significant than fractional improvements from optimized hardware utilization.
I’m a believer that tremendous opportunities are created by hardware innovation. Composable hardware raises interesting questions; however, I get much more excited by open compute (OCP) designs because OCP puts operational concerns first.