07 Multicore Performance

Multicore Processors and System Performance

Multicore Server Architectures

Modern high-performance server scaling relies almost exclusively on multicore processor architectures. To address varying workloads and multi-chip scalability, leading architectures employ distinct organizational approaches for cores, caches, and memory interfaces:

• Intel Xeon Platinum:
  • Utilizes a deeply integrated microarchitecture with up to $40$ cores per chip or up to $224$ cores in a $4$-chip chiplet configuration.
  • Features a distributed $60$ MiB Last Level Cache (LLC) connected by multiple internal routing rings.
  • Operates at slightly lower clock rates than desktop equivalents to remain within stringent thermal power limits.
• IBM Power10:
  • Houses up to $15$ cores per chip, with each core directly coupled to an $8$ MiB L3 cache bank.
  • Connects distributed L3 caches and up to $16$ independent memory channels via $16$ parallel on-chip routing rings.
• AMD EPYC Milan:
  • Utilizes a highly modular chiplet design, incorporating $8$ cores and a shared LLC per chiplet.
  • Standard sockets contain $4$ chiplets, maintaining a bidirectional $32$-byte wide ring interconnect within each chiplet.

The physical layout and cache distribution of these processing cores necessitate specialized topologies for interconnecting multiple chips within a unified system.

Multiprocessor Interconnection Strategies

To scale beyond a single processor, multicore architectures implement sophisticated inter-chip topologies that inherently create Non-Uniform Cache Access (NUCA) and Non-Uniform Memory Access (NUMA) environments:

• IBM Power10 Interconnect:
  • Supports up to $16$ discrete chips ($240$ cores total).
  • Groups $4$ processor chips into a fully connected module via intragroup links.
  • Intergroup links connect each chip to three other modules, ensuring every processor is reachable within one or two network hops.
• Intel Xeon UPI:
  • Uses $3$ Ultra Path Interconnect (UPI) links per processor to interface with neighboring chips.
  • Supports up to $8$ sockets (up to $448$ cores), routed so that any socket is at most two hops away from any other.
• AMD Infinity Fabric:
  • Provides full connectivity among the $4$ to $8$ chiplets residing within a single socket.
  • Dual-socket configurations use remaining Infinity Fabric lanes to bridge the sockets, maintaining the one-to-two hop maximum distance across the entire system.

The efficiency of these interconnection networks directly dictates how well performance scales when executing highly concurrent workloads across hundreds of cores.

Performance Scaling in Multiprogrammed Workloads

Scaling behavior for independent, multiprogrammed tasks depends heavily on the underlying interconnect and memory bandwidth rather than cache coherency overhead. Measured using the SPECintRate benchmark up to $240$ cores, scaling efficiencies diverge significantly among architectures:

• IBM Power10: Exhibits near-linear scalability, achieving $97$ efficiency at $220$ cores relative to its baseline $15$-core configuration.
• Intel Xeon Platinum: Experiences scaling degradation at high core counts, yielding approximately $70$ efficiency at $240$ cores (relative to an $8$-core baseline) and eventually dropping to $50$ efficiency.
• AMD EPYC: Achieves superior raw performance at low core counts but encounters systems-level or architectural limitations as it scales, dropping to $66$ efficiency at $128$ cores relative to a $16$-core baseline.

While independent tasks in multiprogrammed workloads scale primarily based on aggregate memory bandwidth, parallel workloads with active communication exhibit entirely different scaling boundaries.

Workload-Specific Scalability and Energy Efficiency

Workloads requiring shared address space communication behave fundamentally differently than independent request-level workloads when scaled across large NUMA topologies:

• Scientific Parallel Processing (SPEComp2012):
  • Workloads parallelized via OpenMP generate frequent cross-chip communication.
  • On a Xeon Platinum system, speedup remains linear up to $72$ cores, drops to $90$ efficiency between $92$ and $108$ cores, and sharply declines to $\sim 50$ efficiency between $172$ and $220$ cores as the UPI interconnect becomes saturated by coherency traffic.
• Server Processing and Energy Efficiency (SPECpower2008):
  • Models Java server environments containing massive request-level parallelism and minimal inter-process communication.
  • Scales linearly up to $224$ cores without major efficiency drops.
  • Demonstrates strong energy proportionality: at $50$ of maximum load, the system draws only $60$ of its peak power, yielding a relative energy efficiency of $85$ compared to a fully loaded state.

---

Are there specific microarchitectural features of these processors, such as their dynamic scheduling or vector units, that you would like me to detail next?