Confidential 2 The opinions, viewpoints, and any errors or omissions in this presentation are my own. Opinions herein are mine, they do not necessarily reflect those of my employer, Cisco Systems.
Confidential 3 • Systems Architect at Cisco Systems • I work and live in Boulder, CO • Second Best City in the World After Wellington • Concentration on Desktop Virtualization • Data Center Implications - Rich Media • 2 Years at Cisco • Background in Enterprise Applications • Startups, Siebel Systems, Avaya, BOC Gases • Been Writing Thread Reentrant Code for 20+ years • CDC Cyber, VAX 8600, DEC Alpha, x86, ARM
Confidential 10 HDD RPM Advances Stalled at 15k RPM in 2000 Capacity continues to increase, but seek time doesn’t change HDD Experience is 5 years ahead of CPU 1000 5000 10000 15000 x x x x Platter RPM
Confidential 11 • Seagate introduced first 15k RPM HDD in Feb. 2000 Cheetah X15 18.6GB 3.9 ms seek time • 20k RPM HDDs not happening Physics – Heat and Vibration Flash has eliminated market incentives for faster HDDs • Don’t Wait for Multi-TB 15k HDDs either Density and rotational speed don’t play well together
Confidential 13 • Even if Rotational Speed Plateaus, Sequential BW for HDD I/O benefits from improvements in Areal Density • Tightly packed bits pass under the read head faster than loosely packed bits • Random I/O Is pretty well bound to rotational speed • If the bits you need are on other side of platter, then you get to wait • Historically, the Pendulum has swung several times between HDD and CPU technology being the limiting factor for business app performance • Remember HDD Interleave?
Confidential 14 • SANs change the model from managing drives to managing storage • Drives become no more than an component in the SAN • SANs offer improved performance and reliability over individual HDDs • By sequencing writes across multiple platters, IOPS and throughput improve • Think Drum Memory…. Storage Array Disk Drive Simplify The Data Center
Confidential 17 Consolidating services from 20 lightly loaded servers onto a single multicore server makes lots of sense, until that single server crashes for some reason (HW or SW) Consider the prior diagram. Each server probably had a 1Gig Ethernet link, would a single 1 Gig link be sufficient for the whole rack, particularly if the blades don’t have local storage? Compute Density is grand but a big multicore server won’t do you much good if the rest of your environment isn’t equipped to handle it
Confidential 18 It Doesn’t Fix Your Single-Threaded Applications Big Monsters Demand Big Meals MultiCore Servers Require Super-Sized Networking and Storage
Confidential 19 Virtualization Makes it Cheaper and Easier to Create New Virtual Machines for Whatever the OS and Application Combination -or- Virtualization Makes it Easier to Run Amuck With VM Bloat Cisco B440 Full Width Blade 4 Sockets -> 32 Cores 32 DIMM Slots App OS App OS App OS App OS App OS App OS App OS App OS App OS App OS
Confidential 21 • Single Servers Dedicated to Single Services Are Easy • Disaster Recovery and Business Continuity • HW Failure is Coupled To Service Failure • Hosting Multiple Services on a Single Server is Harder • HW Failure Wipes Out Multiplicity of Services • Have to Keep Track of What is Running Where • Load Balancing Takes Thought and Care • Does Little for Storage Issues
Confidential 23 Unified Computing System Server • Unified Computing makes collection of servers the managed entity • No more managing discrete servers • No one should care how many servers they have: Computing matters • Individual HW elements are stateless Simplify The Data Center
Confidential 24 SANs make a bunch of small HDDs look like one big HDD Multicore does not make a bunch of discrete cores look like one big core Multicore is a bit like Nuclear Physics Fission is Easy, Fusion is Hard + == Xeon 5600 Xeon 5600 Xeon 11200 + !=
Confidential 25 Virtualization Web Client Server Minicomputer 2010 2000 1990 1980 1970 1960 First Inflection Point Mainframe Distributed Computing Second Inflection Point
Confidential 29 Cloud Computing Utility Model Workload not VM Centric Plastic Highly Automated Cloud Storage Utility Model Completely Logical Containers Web Service Ifxs High Latency but Parallelizable
Confidential 30 FC Cloud Computing Cloud Storage HTTP Traditional Apps Assume Dedicated HW, OS and App Server connected to High Bandwidth, Low Latency Storage Cloud provides App Container without local persistent storage, abstracted from OS and HW connected to Medium Bandwidth, High Latency Storage through Web Services
Confidential 31 • September 16, 1994 • May 21, 2011 • October 21, 2011 • Harold Camping Numerologist and Failed End of Days Predictor X X X I’m not sure what will come next, but I’ll bet it will seem like déjà vu all over again.