Interesting Links

Links and attachments on various topics related to RAMCloud (most recent additions first):

• SolarFlare performance under VMware ESX
• RocksDB Performance Benchmarks: Reports performance numbers (including latency) for Facebook's key-value store RocksDB
• RAMCube: Exploiting network proximity for RAM-based key-value storage system (RAMCloud).
• Backlight: Machine with very large shared memory (hardware enabled shared coherent memory) (the system at Pittsburgh Supercomputing Center has 32TB) and cores.
• Marvell Dragonfly NVRAM product
• Marvell write acceleration module
• "Research Challenges for Modern Data-Center Networks" by Jeffrey Mogul, HP Labs, 23^rd International Teletraffic Congress (ITC 23) Keynote, referred RAMCloud on p.48, September 8, 2011

• "Warehouse-Scale Computing: Entering the Teenage Decade" by Luiz Barroso, Google, Federated Computing Research Conference  (FCRC) Plenary, June 8, 2011 (referred RAMCloud as a notable technology around 45min. in the video)

• The Amazon EC2 meltdown in April 2011
• Performance of memcached and Hadoop over Infiniband (Slides from OpenFabrics Monterey Workshop, April 2011).
• Other large social graphs
• Estimate of the size of the Twitter social graph
• Introduction to Performance Analysis on Nehalem Based Processors
• AGIGATech non-volatile memory DDR3 memory with on-board Flash and supercap; automatically backs itself up to flash after a power failure
• Foursquare meltdown in October 2010
• Facebook meltdown in September 2010: Bobby Johnson's analysis, Todd Hoff's discussion.
• ChronicDB: commercial system that allows database schemas to be updated without stopping applications
• Bitcask storage system
• NoSQL benchmark results from Yahoo!
• Hash table performance measurements (from incise.org)
• Facebook statistics
• Redis: a persistent main-memory key-value store; looks like it has pretty good performance.
• High Scalability blog article: written in March 2009 by Todd Hoff, covers a lot of the RAMCloud ideas and mentions several commercial products claimed to be shipping with memory-based solutions. Here are some additional Web pages that are related to this blog entry:
  • http://highscalability.com/handle-1-billion-events-day-using-memory-grid
  • http://davidvancouvering.blogspot.com/2009/02/why-not-just-put-whole-friggin-thing-in.html
  • http://gevaperry.typepad.com/main/2009/03/cloud-pricing-and-application-architecture.html
• Technical report on using MapReduce on top of memcached: Low-Latency, High-Throughput Access to Static Global Resources within the Hadoop Framework
• Talk by Andy Bechtolsheim on storage technologies
• Arista Web site: for their 48-port switches, latency is 0.9 µs.
• "Drop ACID and think about data" (talk by Bob Ippolito): Video, Commentary/Summary
• http://www.ddrdrive.com/ - an interesting block storage device implemented as 4 sticks of DDR (backed by flash) on a PCIe x1 board. 300000 IOPS (512 byte random reads and writes).
• Memcached benchmarks on Nehalem - Sun benchmarks memcached, running on the latest Nehalem processors over 10G links (they get > 350k ops per second, with ~ 62% network and 52% CPU utilization)
• Powerpoint slides from the InfoLunch discussion of RAMCloud on April 10, 2009.
• Slides from IBM on the role of non-volatile memory.
• Technical overview of Facebook Haystack photo repository
• Gordon: Peta-scale Solid-state Storage Systems
• The FAWN project at CMU: optimizing queries/joule using small nodes.
• You-tube video on Google datacenters