Addressing

Some possibilities

• Structured, Unstructured
• Random, Hashes, Sequential
• User-specified, generated
• Need at least 2^48 capacity for objects
  • Hence, unstructured addresses probably need to be at least 2^64
  • 64 * 2^30 bytes/machine * 2^14 machines = 2^50 bytes, 2^50 bytes/2^7 bytes/obj = 2^48 objects

Sequential and Structured

• Temporal id locality
• Allocation could be tricky to make fast
• How many tables does a typical (or large) RDBMS have?
• How many applications do we expect to support in a single RAMCloud instance?
• How much metadata space is need for all tables/applications?
• How does metadata replication occur and what is the frequency?

Random

• Smaller ids (64-bit?)
  • Not if we want these to look like capabilities
• Simple to make fast
• Not meaningful to client (both a plus and minus)
• Indexing must be done by clients and stored in the cloud
  • Akin to FriendFeed's setup

Distribution

Effects

• + Capacity
  • Strictly necessary for this reason alone
• + Throughput
  • + Avoid hot spots
• - Consistency

• How do durability and reliability tie in to the discussion?

Approaches

Mapping

Tradeoff: Capacity, Throughput (via parallel requests) vs Latency (lookup time client->host, host->addr+metadata)

RAMCloud Address -> Metadata as quickly as possible (e.g. with as few requests and as little processing as possible) where metadata includes physical address, size, and permissions at least.

Ideal: 0 network messages and O(1) address to host mapping time with high probability

Implies all clients are aware of mapping.
Complication: access-control requires highly-consistent mapping replication if control is on addresses (e.g. the application/table is part of the structured address).

Objects may need to relocate due failures, load, or capacity.

DHT

• + Simple
• + Natural replication
• - Latency
  • Address to shard mapping has log(# shards) time in general
  • Can be mitgated for index space tradeoff using radix tree or tries
  • How many levels is too deep? Even 2-3 in the face of cache misses?
• + Load sharing
• - More difficult to co-locate related data on a single machine
  • Probably the case that we want to intentionally distrbute related data (more network overhead, but reduces latency because lookups happen on independent machines)

• Extensible
• Linear
• Consistent

B-Trees

• + Supports range queries on totally ordered keys
• +/- Allows several records from the same table to be returned with a single request
• May cause a server to become a hot spot
  • Is this anymore true than with hashing?

Replication

Effects

• + Throughput
  • + Mitgates hot spots
  • If there is a hot spot, data reduction may not solve overloading problems
  • If a single server can handle 1M requests/second, is there any need to replicate?
  • If the load gets too high, perhaps reduce the load by reducing the amount of data stored on a server, rather than replicating the data
• + Latency
  • Eliminates cross data center requests (East to West Coast datacenters)
• - Consistency
  • A system without replication would be much easier to manage

Locality

Effects

• + Network traffic
• + Latency for serial requests
• + Performance isolation in multi-tenant environments
• + Economy of metadata
  • For example, only access control information for data which resides on a host must be replicated to that host

• Is there any locality in interesting database applications?
• The most interesting form of locality is locality within a request: would like to satisfy each request with a single call to a single server, if possible