• Underlying network in datacenter has become so fast (RDMA, CXL, etc.) that the latency of accessing remote server (from Server A → Server B) is the same as latency of accessing local server
• Idea is to split the servers and connect them using high performance network
  • CPU-heavy rack
  • Memory-heavy rack
  • Persistent storage rack
  • etc.
• What should OS do? What should orchestrating services like Kubernetes do?
• Pros of disaggregated datacenters
  • Easy to scale out/in, easier to scale out individual things, like if you’re good on memory but need more compute power, just expand your CPU servers
• Latency increases with memory size! When you access a memory address, translation takes a long time as it needs to go through a set of registers that increases in number as memory itself scales

Erasure Coding

• Comes from communications and has been adopted by distributed systems
• Alternative to replication in fault-tolerance is to store parity blocks, XOR them to recover data
• Now erasure encoding builds on this
  • packet split into D packets in such a way that
  • any K subset of the D packets is sufficient to decode and retrieve the original packet
  • ex: if packet split into D = 8 packets, we only need K = 5 servers to respond to reconstruct the original packet, so only the first 5 responses are sufficient

SWARM and ACESO (read in class) are both specialized for KV-stores, but consensus based replication is still an open problem for transactions? Indy says we can expect papers in this area soon!