230 likes | 380 Views
LFTI: A Performance Metric for Assessing Interconnect topology and routing design. Background Innovations in interconnect topology and routing design is essential for future generation ultra-scale supercomputers. Current methods for evaluating topology and routing design are not ideal.
E N D
LFTI: A Performance Metric for Assessing Interconnect topology and routing design • Background • Innovations in interconnect topology and routing design is essential for future generation ultra-scale supercomputers. • Current methods for evaluating topology and routing design are not ideal.
Current methods for evaluating interconnect topology and routing design • Topology and routing are evaluated separately • Topology • Diameter, bisection bandwidth, nodal degree, etc • Not directly related to application level performance • Routing with topology • Simulation to get throughput and packet latency • Limited network sizes and numbers of scenarios • Simulation sees the tree, but not the forest. • Two kinds of metrics: simple metrics that do not directly relate to performance and detailed metrics that are too expensive to obtain.
Impact of evaluation methods • Evaluation methods set the design optimization objective • Recently proposals (dragonfly, jellyfish) all have large bisection bandwidth and support certain traffic patterns effectively. • Think of how the designs are justified!! • Excellently designs with traditional metrics. • Are these designs good for typical HPC workloads? • There is no metric that can be used to compare across different topology and routing designs for HPC workloads.
What kind of metrics are we looking for? • Desirable properties: • Reflect overall network performance • Simple enough that it can be computed quickly – we do not want to do simulation. • A related attempt -- effective bisection bandwidth: summarize network performance by the average performance for all bisection communication patterns. • Is this metric reflective?
LFTI: LANL-FSU throughput indices • A metric for throughput performance • High level ideas • Use modeling the obtain the average throughput for one communication pattern. • Find the set of representative communication patterns to be used in the metrics • Summary the overall network performance using the average throughput performance for a large number of communication patterns common to HPC applications
LFTI: LANL-FSU throughput indices • High level ideas • Once the patterns to be included is determined, LFTI can be derived from most topology and routing specifications without detailed simulation. • If an interconnect can achieve high overall performance for many common HPC patterns, it is likely that it will provide high performance for HPC workloads. • Unlike some other metrics, LFTI is much harder to cheat.
LFTI: LANL-FSU throughput index • LFTI is the summary of the throughput of an interconnect for a large number of common communication patterns in HPC applications. • For each communication pattern, a metric (sustained throughput) is used that is closely related to the application level performance for that pattern to quantify the performance of the interconnect. • For a class of patterns (e.g. 2DNN patterns), the expected sustained throughput is used to quantify the performance. • LFTI is the aggregate of the performance of many classes of patterns.
Computing the sustained throughput for a pattern (single path routing) • Compute the link load (number of flows going through each link) • The sustained throughput for each flow is its share of the throughput on the bottleneck link or Max-Min fairness. • The sustained throughput for the pattern is the aggregate throughput of all flows in the pattern. • Normalized with per flow throughput divided by the input link bandwidth.
Computing the throughput index for a class of patterns • A throughput index for a class of patterns (e.g. 2DNN patterns) is the expected sustained throughput across all patterns of that class. • The index can be obtained by randomly sampling of a large number of patterns (e.g. 10000 patterns) • May apply some statistical method to obtain the index with confidence without sampling a large number of patterns.
Communication Patterns in LFTI indices • Patterns with history • All to all, • Bisect – effective bisection bandwidth • Low-dimensional stencil patterns • 2DNN, 2DNN_DIAG, 3DNN, 3DNN_DIAG • Random patterns – for applications with unstructure mesh, adaptive mesh refinement methods • RANDOM 50, RANDOM N50 • Commonly used sub-communication patterns • Permutation, shift
LFTI categories • Trying to reflect how the machine is used • Whole system direct map LFTI • Whole system random map LFTI • Job allocation trace-based LFTI • Largest job based on some job traces
Evaluating interconnect using LFTI Fat-tree (ftree), dragonfly (dfly), hypercube(hcube) 6D torus (6D), 3D torus (3D), jellyfish (jfish) of 25K-35K nodes – the size of the next generation supercomputer.
Conclusion • Traditional performance metrics such as bisection bandwidth and effective bisection bandwidth are not indicative for interconnect’s performance. • Optimizing for BB and EBB may not lead to high performance interconnects. • LFTI is indicative of application level performance, yet can be derived rapidly without detailed simulation. • It is a much better metric than the current metrics.
LFTI weakness • Communication patterns and weights • Heavily concentrating on simulation types of applications • Not much for data intensive applications • Calls for performance characterization work • To find the truly “representative” workload to be included in the index.
LFTI weakness • LFTI relies on fast modeling of throughput performance from each communication patterns • Depending on the routing algorithm, the modeling can be problematic • Indirect adaptive routing is an example – no effective model method than simulation. • Needs to develop new models for all existing and future routing schemes, and whatever can affect the “sustained throughput”