HP will start delivering memristor-based RAM DIMMs in 2016, and the Machine itself is expected to be available as a product by 2019. But within the next year, HP will release an open-source Machine OS software developer’s kit and start producing prototypes for collaboration with software vendors.
Some enterprise systems already use nonvolatile RAM based on battery-powered DRAM to help prevent data loss in the event of a power outage. But memristor RAM is an entirely different thing—it could theoretically allow for computers to start processing again in the exact same state they were in before they were disconnected from power. That would make “instant on” devices much more power-efficient, but it would also completely change how operating systems deal with system resets and powering down—they would have to figure out what needed to be kept in memory and what needed to be cleared before restarting. Errors could result if certain areas in memory weren’t cleared.
Sontag said that programming languages will also have to change. “In the very long term,” Sontag said, “we have to change the memory semantics of programming languages to make it possible to say what is stored in nonvolatile memory and what isn’t. We need to come to an agreement with the industry about the semantics for nonvolatile RAM.”
As part of an effort to create that agreement, HP is turning to the open source community. “We have a number of approaches for what stays in nonvolatile memory that we’ll sort through in the next year and then take it to the Linux community,” said Sontag.
* low latency access to universal memory
* orders of magnitude more energy efficient
* special purpose cores focuses on the compute that you are doing – customize hardware to change cost energy curve
Any byte in 160 petabytes in 160 racks can be addressed in less than 250 nanoseconds [enabled by photonics]
Memory Cube photonically connected to the processor
3Par memory management is key for universal memory management.
Rejuvenate operating system research for universal memory based systems. Do not need to spend 90% of the effort moving data from storage to memory.
HP is making linux and android versions tuned for non-volatile memory.
The memristor memory cloud
The first target of opportunity for memristor RAM, however, won’t require a total overhaul of computing—large in-memory data stores that could take the place of solid-state disk storage and existing in-memory data stores.
“We already have in-memory disks and filesystems that we emulate,” Sontag said. “They use DRAM and just happen to be volatile. You can point an in-memory virtual disk at memristor, and now it is nonvolatile—it preserves the semantics of all the software running it. Software still sees it as a block device—just much quicker response time. Then we’ll work our way into block stores and databases, optimized for this space—flatten their memory hierarchy, and manage how they allocate store in nonvolatile memory. “
That could lead to a changes in how cloud giants like Google and Facebook develop their software—making it a lot less complicated. For example, memristor memory could help remove some of the hassle of dealing with large structured and unstructured databases—particularly the parts that revolve around keeping just the right information in cache. “The challenge with today’s technology is that caches are good if they have a good hitrate,” Sontag said, “but as soon as your workflow starts to change, the cache becomes more and more inefficient.”
Sontag envisions memristor memory eventually finding its way throughout the whole cloud—moving data closer to the endpoint by putting memory caches similar to those used by services like Cloudflare and down to the endpoints themselves.