RAMCloud and the Low-Latency Datacenter

RAMCloud and the Low-Latency Datacenter

John Ousterhout, Stanford

Low-latency Datacenter in General:

     Datacenter: large compute power + large amount of storage + fast communication, homogeneous, centrally controlled

      How many datacenters: $1K-10K per person for IT infrastructure?

      Innovations in Datacenters:

               Phase 1: manage scale (map-reduce)

                Phase 2: low latency (from 1-2ms to 2-10 us)

                                 In current datacenter, application is separated from its data (unlike in a single machine)

RAMCloud (Storage system for low latency datacenter)

Goal:

1.       All data in DRAM (no cache misses)

2.       Durable and available

3.       Large scale

4.       Low latency (5-10us remote access) for small chunks of data

Architecure:

                App + network + storage servers + coordinator

                Storage servers: master + backup. Master manages memory and data requests, backup manages secondary storage and durability, talking to other masters for availability etc.

               

Operations:

                Read, write, delete, cwrite(conditional write based on version number)

Data Durability:

                Eventual durability, log structure approach (with small non-violtial memory).

                Crash recovery: log replaying in large scale (?)

Log-structured memory:

            Don’t use malloc (waste at least 50% memory because of workload change), instead, structure memory as a log (80%-90% memory utilization)

New Projects:

                Higher level data model  (for database, say, secondary indexes?)
                System architecture for RPC

               

Threats to Latency:

                Layering (how to organize software structure but still low latency?)

                Network buffering (esp. TCP just tries to find empty buffers!)