Building a Beowulf: My Perspective and Experience Ron Choy Lab. for Computer Science MIT Outline • • • • History/Introduction Hardware aspects Software aspects Our class Beowulf The Beginning • Thomas Sterling and Donald Becker CESDIS, Goddard Space Flight Center, Greenbelt, MD • Summer 1994: built an experimental cluster • Called their cluster Beowulf The First Beowulf • • • • 16 x 486DX4, 100MHz processors 16MB of RAM each, 256MB in total Channel bonded Ethernet (2 x 10Mbps) Not that different from our Beowulf The First Beowulf (2) Current Beowulfs • Faster processors, faster interconnect, but the idea remains the same • Cluster database: http://clusters.top500.org/db/Query.php3 • Top cluster: 1.433 TFLOPS peak Current Beowulfs (2) What is a Beowulf ? • Massively parallel computer built out of COTS • Runs a free operating system (not Wolfpack, MSCS) • Connected by high speed interconnect • Compute nodes are dedicated (not Network of Workstations) Why Beowulf? • It’s cheap! • Our Beowulf, 18 processors, 9GB RAM: $15000 • A Sun Enterprise 250 Server, 2 processors, 2GB RAM: $16000 • Everything in a Beowulf is open-source and open standard - easier to manage/upgrade Essential Components of a Beowulf • • • • Processors Memory Interconnect Software Processors • Major vendors: AMD, Intel • AMD: Athlon MP • Intel: Pentium 4 Comparisons • Athlon MPs have more FPUs (3) and higher peak FLOP rate • P4 with highest clock rate (2.2GHz) beats out the Athlon MP with highest clock rate (1.733GHz) in real FLOP rate • Athlon MPs have higher real FLOP per dollar, hence it is more popular Comparisons (2) • P4 supports SSE2 instruction set, which perform SIMD operations on double precision data (2 x 64-bit) • Athlon MP supports only SSE, for single precision data (4 x 32-bit) Memory • DDR RAM (double data rate) – used mainly by Athlons, P4 can use them as well • RDRAM (Rambus DRAM) – used by P4s Memory Bandwidth • Good summary: http://www6.tomshardware.com/mainboard/ 02q1/020311/sis645dx-03.html • DDR beats out RDRAM in bandwidth, and is also cheaper Interconnect • The most important component • Factors to consider – – – – Bandwidth Latency Price Software support Ethernet • Relatively inexpensive, reasonably fast and very popular. • Developed by Bob Metcalfe and D.R. Boggs at Xerox PARC • A variety of flavors (10Mbps, 100Mbps, 1Gbps) Pictures of Ethernet Devices Myrinet • Developed by Myricom • “OS bypass”, the network card talks directly to host processes • Proprietary, but very popular because of its low latency and high bandwidth • Usually used in high-end clusters Myrinet pictures Comparison Latency Fast Ethernet ~120s Gigabit Ethernet ~120 s Bandwidth ~100Mbps ~1Gbps peak peak Myrinet ~7 s ~1.98Gbps real Cost Comparison • To equip our Beowulf with: – Fast ethernet: ~$1700 – Gigabit ethernet: ~ $5600 – Myrinet: ~$17300 How to choose? • Depends on your application! • Requires really low latency e.g. QCD? Myrinet • Requires high bandwidth and can live with higher latency e.g. ScaLAPACK? Gigabit ethernet • Embarrassingly parallel? Anything What would you gain from fast interconnect? • Our cluster: Single fast ethernet (100Mbps) – 36.8 GFLOPS peak, HPL: ~12 GFLOPS – 32.6% efficiency • GALAXY: Gigabit ethernet – 20 GFLOPS peak, HPL: ~7 GFLOPS – 35% efficiency *old, slow tcp/ip stack!* • HELICS: Myrinet 2000 – 1.4 TFLOPS peak, HPL: ~864 GFLOPS – 61.7% efficiency My experience with hardware • How long did it take for me to assemble the 9 machines? 8 hours, nonstop Real issue 1 - space • Getting a Beowulf is great, but do you have the space to put it? • Often space is at a premium, and Beowulf is not as dense as traditional supercomputers • Rackmount? Extra cost! e.g. cabinet ~$1500, case for one node ~$400 Real issue 2 – heat management • The nodes, with all the high powered processors and network cards, run hot • Especially true for Athlons - can reach 60°C • If not properly managed, the heat can cause crash or even hardware damage! • Heatsink/fans – remember to put in in the right direction Real issue 3 - power • Do you have enough power in your room? • UPS? Surge protection? • You don’t want a thunderstorm to fry your Beowulf! • For our case we have a managed machine room - lucky Real issue 4 - noise • Beowulfs are loud. Really loud. • You don’t want it on your desktop. Bad idea Real issue 5 - cables • Color scheme your cables! Software • We’ll concentrate on the cluster management core • Three choices: – Vanilla Linux/FreeBSD – Free cluster management software (a very patched up Linux) – Commercial cluster management software (very very patched up Linux) The issues • Beowulfs can get very large (100’s of nodes) • Compute nodes should setup themselves automatically • Software updates must be automated across all the nodes • Software coherency is an issue Vanilla Linux • Most customizable, easiest to make changes • Easiest to patch • Harder for someone else to inherit the cluster – a real issue • Need to know a lot about Linux to properly setup Free cluster management softwares • • • • Oscar: http://oscar.sourceforge.net/ Rocks: http://rocks.npaci.edu MOSIX: http://www.mosix.org/ (usually patched) Linux that comes with software for cluster management • Reduces dramatically the time needed to get things up and running • Open source, but if something breaks, you have one more piece of software to hack Commercial cluster management • Scyld: www.scyld.com - founded by Donald Becker • Scyld – father of Beowulf • Sells a heavily patched Linux distribution for clustering, free version available but old • Based on bProc, which is similar to MOSIX My experience/opinions • I chose Rocks because I needed the Beowulf up fast, and it’s the first cluster management software I came across • It was a breeze to setup • But now the pain begins … severe lack of documentations • I will reinstall everything after the semester is over Software (cont’d) • Note that I skipped a lot of details: e.g. file system choice (NFS? PVFS?), MPI choice (MPICH? LAM?), libraries to install … • I could talk forever about Beowulfs but it won’t fit in one lecture Recipe we used for our Beowulf • Ingredients: $15000, 3 x 6 packs of coke, 1 grad student 1. Web surf for 1 weeks, try to focus on the Beowulf sites, decide on hardware 2. Spend 2 days filling in various forms for purchasing and obtaining “competitive quotes” 3. Wait 5 days for hardware to arrive, meanwhile web surf some more, and enjoy the last few days of free time in a while Recipe (cont’d) 4. Lock grad student, hardware (not money), and coke in an office. Ignore scream. The hardware should be ready after 8 hours. Office of the future Recipe (cont’d 2) 5. Move grad student and hardware to its final destination. By this time grad student will be emotionally attached to the hardware. This is normal. Have grad student set up software. This would take 2 weeks. Our Beowulf Things I would have done differently • No Rocks, try Oscar or may be vanilla Linux • Color scheme the cables! • Try a diskless setup (saves on cost) • Get rackmount Design a $30000 Beowulf • One node (2 processors, 1GB RAM) costs $1400, with 4.6 GFLOPS peak • Should we get: – 16 nodes, with fast ethernet, or – 8 nodes, with Myrinet? Design (cont’d) • 16 nodes with fast ethernet: – 73.6 GFLOPS peak – 23.99 GFLOPS real (using the efficiency of our cluster) – 16 GB of RAM • 8 nodes with Myrinet – 36.8 GFLOPS peak – 22.7 GFLOPS real (using the efficiency of HELICS) – 8 GB of RAM Design (cont’d 2) • First choice is good if you work on linear algebra applications, and require lots of memory • Second choice is more general purpose