1. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 1 RAMCloud Overview ● Storage for datacenters ● 1000-10000 commodity servers ● 32-64 GB DRAM/server ● All data always in RAM ● Durable and available ● Performance goals:  High throughput: 1M ops/sec/server  Low-latency access: 5-10µs RPC Application Servers Storage Servers Datacenter

2. Example Configurations For $100-200K today:  One year of Amazon customer orders  One year of United flight reservations CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 2 Today5-10 years # servers20004000 GB/server24GB256GB Total capacity48TB1PB Total server cost$3.1M$6M $/GB$65$6

3. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 3 RAMCloud Motivation: Latency ● Large-scale apps struggle with high latency  Facebook: can only make 100-150 internal requests per page UI App. Logic Data Structures Traditional Application UI Bus. Logic Application Servers Storage Servers Web Application << 1µs latency0.5-10ms latency Single machine Datacenter

4. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 4 RAMCloud Motivation: Latency ● RAMCloud goal: large scale and low latency ● Enable new class of applications:  Crowd-level collaboration  Large-scale graph algorithms UI App. Logic Data Structures Traditional Application UI Bus. Logic Application Servers Storage Servers Web Application << 1µs latency 0.5-10ms latency Single machine Datacenter 5-10µs

5. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 5 RAMCloud Motivation: Technology Disk access rate not keeping up with capacity: ● Disks must become more archival ● More information must move to memory Mid-1980’s2009Change Disk capacity30 MB500 GB16667x Max. transfer rate2 MB/s100 MB/s50x Latency (seek & rotate)20 ms10 ms2x Capacity/bandwidth (large blocks) 15 s5000 s333x Capacity/bandwidth (1KB blocks) 600 s58 days8333x

6. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 6 RAMCloud Research Issues ● Data durability/availability ● Fast RPCs ● Data model, concurrency/consistency model ● Data distribution, scaling ● Automated management ● Multi-tenancy ● Client-server functional distribution ● Node architecture

7. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 7 Data Durability/Availability ● Data must be durable and available when write RPC returns ● Unattractive approaches:  Replicate in other memories (too expensive)  Synchronous disk write (100-1000x too slow) ● Our approach: buffered logging DRAM disk DRAM disk Storage Servers write log async, batch DRAM disk log

8. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 8 Buffered Logging,cont’d ● Potential problem: power loss  Per-server battery backup?  Nonvolatile memory on disk controllers? ● Potential problem: crash recovery  If master crashes, data unavailable until recovered from disks on backups  Read 64 GB from one disk? 10 minutes  Our goal: recover in 1-2 seconds ● Solution: take advantage of system scale  Scatter backup data across many servers  Recover in parallel

9. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 9 Recovery, First Try ● Scatter log segments randomly across all servers ● After crash, all backups read disks in parallel (64 GB/1000 backups @ 100 MB/sec = 0.6 sec) ● Collect all backup data on replacement master (64 GB/1GB/sec ~ 60 sec: too slow!) Backups... Replacement Master

10. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 10 Recovery, Second Try ● Divide each master's data into partitions ● Recover each partition on a separate server:  100 partitions, 640 Mbytes each  1 GB/sec NIC per replacement master  Recovery time < 1 sec Backups... Replacement Masters Dead Master

11. CS 142 Lecture Notes: Large-Scale Web ApplicationsSlide 11