Transcript Slide 1
SEDCL: Stanford Experimental Data Center Laboratory Tackle Data Center Scaling Challenges with Stanford’s research depth and breadth Data Center Scaling • A network of data centers and web services are the key building blocks for future computing • Factors contributing to data center scaling challenges – Explosive growth of data with no locality of any kind – Legal requirement to backup data in geographically-separated locations---big concern for financial industry – Emergence of mobile and Cloud Computing – Massive “interactive” web application – Energy as a major new factor and constraint – Increasing capex and opex pressures • Continued innovations critical to sustain growth 3 Stanford Research Themes • RAMCloud: main-memory based persistent storage – Extremely low latency RPC • Networking: – Large, high-bandwidth, low-latency network fabric – Scalable, error-free packet transport – Software defined data center networking with OpenFlow • Servers and computing – Error and failure resilient design – Energy aware and energy proportional design – Virtualization and mobile VMs 4 Major research topics of SEDCL • RAMCloud: Scalable DRAM-based Storage – Scalable nvRAM – All data in DRAMs all the time • Interconnect fabric – Bufferless networks: low-latency, high-bandwidth network • Packet transport – – – – • Reliable delivery of packets: R2D2—L2.5 Congestion management: QCN (IEEE 802.1Qau), ECN-HAT, DCTCP Programmable bandwidth partitioning for multi-tenanted DCs: AF-QCN Low-latency 10GBaseT Related projects – OpenFlow – Energy aware and energy proportional design 5 Experimentation is Key to Success • Many promising ideas and technologies – Will need iterative evaluation at scale with real applications • Interactions of subsystems and mechanisms not clear – Experimentation best way to understand the interactions • Difficult to experiment with internal mechanisms of a DC – No experimental facilities and that is a big barrier to innovations • Ongoing efforts to enable experimentation – Facebook, Microsoft, NEC, Yahoo!, Google, Cisco, Intel, … 6 Overview of Research Projects • RAMCloud • Packet transport mechanisms – Reliable and reliable data delivery: R2D2—L2.5 – ECN-HAT, DCTCP: collaboration with Microsoft • Data center switching fabric – Extremely low latency, low errors and congestion (bufferless) – High port density with very large bisection bandwidth project just initiated 7 RAMCloud Overview Lead: John Ousterhout • Storage for datacenters • 1000-10000 commodity servers • 64 GB DRAM/server • All data always in RAM • Durable and available • Low-latency access: 5µs RPC • High throughput: 1M ops/sec/server Application Servers Storage Servers Datacenter RAMCloud Research Issues • Data durability and availability • Low latency RPC: 5 microseconds – Need suitable network! • Data model • Concurrency/consistency model • Data distribution, scaling • Automated management • Multi-tenancy • Client-server functional distribution Layer 2.5: Motivation and use cases • Speed up TCP performance in data centers • TCP performs poorly when there is a large number of packet drops • Applications like MapReduce/Hadoop and GFS cause the “incast problem” where a large number of packets are dropped at switches • FCoE • Corruption losses, though rare, lead to SCSI timeouts. • Priority flow control (IEEE 802.1Qbb) enables Ethernet switches not to drop packets, but requires “skid” or “PAUSE-absorption” buffers. But skid buffers grow as bandwidth-delay product of links and are very expensive. 10 L2.5 Research Issues • Determine simple signaling method – Simplify (or get rid of) headers/tags for L2.5 encapsulation • Develop and refine the basic algorithm for TCP – In the kernel – In hardware (NICs) • Develop the algorithm for storage (FC, FCoE) • Deploy in a large testbed • Collaborate on standardization DCTCP • DCTCP: TCP for data centers – Operates with really small buffers – Optimized for low-latency – Uses ECN marking with Mohammad Alizadeh, and Greenberg et al at Microsoft Influenced by ECN-HAT (with Abdul Kabbani) DCTCP: Transport Optimized for Data Centers 2. 3. High throughput – Creating multi-bit feedback at TCP sources Low Latency (milliseconds matter) – Small buffer occupancies due to early and aggressive ECN marking Burst tolerance – Sources react before packets are dropped – Large buffer headroom for bursts 157 160 DCTCP Completion Time (ms) 1. TCP DeepBuf 145 RED 120 86 80 Queue buildup 29 40 15 15 Incast 24 21 0 Short messages 1. 2. Query Sauce Use full info in stream of ECN marks Adapt quickly and in proportion to level of congestion Mark Packet buffer K Don’t Mark DCTCP ECN Marks DCTCP TCP 1011110111 Cut window by 40% Cut window by 50% 0000000001 Cut window by 5% Cut window by 50% Reduces variability Reduces queuing Research Themes and Teams WEB App Framework Resilient Systems Virtualization: S. Mitra N. McKeown M. Rosenblum B. Prabhakar K. Kozyrakis Energy Aware P. Levis N. McKeown J. Ousterhout Storage M. Rosenblum D. Mazieres Server and network Networking 14 N. McKeown G. Parulkar B. Prabhakar J. Ousterhout M. Rosenblum