PLNOG15 -The Exascale Future of Today (Rob Bird)

### Overview Robert Bird, head of Big Data Architecture and Big Data Research at Akamai Technologies, examines the state and challenges of exascale computing (processing quintillions of operations, bytes, or bits per second). Drawing from Akamai's operational reality—delivering 30% to 50% of global web traffic, peaking at 40 Tbps across roughly 250,000 servers, and managing about 5 petabytes of logs per day—Bird assesses why the anticipated arrival of exascale computing has stalled since predictions in 2007. The presentation breaks down exascale feasibility across three primary pillars: compute architectures, networking, and storage, while emphasizing the convergence between traditional supercomputing, cloud architectures, and peer-to-peer (P2P) systems.

### Key Topics - **Defining Exascale:** Exascale refers to quintillion-scale operations (exaflops), data storage (exabytes), or transmission rates. While storage has already entered the exascale tier, compute is hovering near petascale thresholds, and true exascale networking remains largely theoretical. - **Real-World Drivers:** Projects driving exascale demand include high-energy physics (the Large Hadron Collider, which could produce 36 exabytes per day but limits retention to 30 petabytes per year due to transmission constraints), astrophysics (the Square Kilometre Array), 8K video streaming, intelligence cartography, and large consumer web platforms like Akamai, Google, Facebook, and Netflix. - **The Networking Bottleneck:** Transmitting an exabyte over a 100 Gbps link would require roughly 1,000 days (or 2.5 days across Akamai’s entire global bisection bandwidth), highlighting data movement as the single most prohibitive factor in exascale progression. - **Storage, Multi-Tenancy, and Data Locality:** Effective big data storage requires moving past simple bulk storage (e.g., standard HDFS) toward anticipatory caching, multi-tiered data placement, and database architectures capable of handling heterogeneous queries.

### Technical Details - **Compute and Efficiency Limits:** Achieving an exaflop using smartphone-class hardware (e.g., ARM Cortex-A9 in iPhone 6) would take one billion units consuming roughly 7 gigawatts. Leading supercomputers, such as Tianhe-2 (achieving roughly 35 to 55 petaflops peak at 18 to 30+ megawatts), demonstrate extreme power requirements. In contrast, purpose-built networks like Bitcoin reach 5.4 exahashes/sec via ASICs, and the human brain operates at an estimated exaflop-equivalent pattern recognition capability on just 20 watts at 10 Hz. - **Interconnects and Signaling:** High-performance computing utilizes custom fabrics (e.g., Cray Titan, IBM Blue Gene) and InfiniBand, which reaches near 1 Tbps aggregate bandwidth and is evolving into routable datacenter environments. Optical signaling approaches theoretical limits around 1.2 petabits/sec; an experimental 255 Tbps fiber link demonstrated in the Netherlands only spanned approximately two feet. - **Storage Formatting and Query Processing:** Big data architectures face significant trade-offs between write efficiency and read optimization. Once multi-exabyte stores are committed in specific formats (row vs. column vs. pre-aggregated cubes), transforming them is computationally prohibitive. Distributed database systems like Google Spanner face graph-based Hamiltonian path optimizations for query planning across mobile users, pointing toward Michael Stonebraker's concept of "polystores"—multi-engine platforms capable of concurrently evaluating relational, graph, full-text, and array data.

### Takeaways - Moore's Law scaling based on raw clock speed and transistor doubling has plateaued, shifting the design focus toward non-von Neumann models, neuromorphic computing, FPGAs integrated into server chips, and specialized accelerators (GPUs/ASICs). - Exascale datacenters will require radical physical design shifts, including floor-based passive immersion cooling handling 200 kW to 1 MW per rack. - Progress depends on cross-pollination: supercomputing contributes insights into power management, interconnect hierarchies, and MPI-style parallelization, while cloud computing provides lessons in commodity resilience, planetary caching hierarchies, and software-defined architectures.