Required Readings⌗

Optional⌗

Summary⌗

Fall off of Moore’s law leads to specialization and heterogeneity compute

  • GPUs
  • TPUs
  • New Memory/Storage classes

Disaggregation := independently scaled tiers of different resources

What is RDMA⌗

Remote Direct Memory Access

Bypass CPU involvement in data access via interconnect

Higher bandwidth and lower latency, but higher cost per port

Two sided RDMA

  • Traditional send/recv semantics like sockets

One sided RDMA

  • CPU access remote memory directly
  • Memory needs to be pinned and not swapped out to disk

RDMA Specialized RPC⌗

One sided RDMA is faster, but needs a redesign

  • Plus RPC typically has a service that needs to be accessed anyway

Instead we can create a new type of RPC that leverages RDMA features

  • Connectionless protocols
  • Shared receive queues

What If Memory Is Persistent⌗

Intel Optane was Byte-Addressable Persistent Memory (PMEM)

Persistent data operations require flush to persistent memory

  • Must complete before client is acknowledged
  • Removes advantage of RDMA over send/recv RPC

Disaggregation⌗

Pools of network attached but independently scaled resources

Not new, but easier with faster interconnects and smarter devices

Disaggregating CPU and Memory with LegoOS⌗

Colocate Virtual Memory System onto MMU instead of CPU

Cache misses now have to go over the network which is much slower and latent

  • Help this by adding large CPU “Extended Cache”

LegoOS Select Experimental Result⌗

Prototype implemented in emulation

Monolithic servers, but all but some resources ignored

  • So “Network attached Hard drives” were regular servers that didn’t utilize CPU, RAM, etc

Controllers implemented in Linux

Connected via RDMA network, communicated via RPC

Actual system designs:

  • HPE “The Machine”
  • Interconnects for Fabric Attached Memory (OpenFAM)
  • Berkeley Firebox system

Baseline Comparisons:

  • Linux with SSD Swap
  • Linux with Ramdisk Swap
  • Linux with InfiniSwap
  • LegoOS

Workloads:

  • Unmodified TensorFlow running CIFAR-10
    • Working set: 0.9G
    • 4 threads

LegoOS was much better