Bull eXascale Interconnect
Exascale entails an explosion of performance, of the number of nodes/cores, of data volume and data movement. At such a scale, optimizing the network that is the backbone of the system becomes a major contributor to global performance. The interconnect is going to be a key enabling technology for exascale systems. This is why one of the cornerstones of Bull’s exascale program is the development of our own new-generation interconnect.
The Bull eXascale Interconnect or BXI introduces a paradigm shift in terms of performance, scalability, efficiency, reliability and quality of service for extreme workloads.
The BXI fabric is highly scalable (up to 64.000 nodes for the first version), it features:
- High-speed links (100 Gb/s/s)
- High message rate (>100 M msg/s)
- Minimal memory footprint and low latency components
Getting rid of the communications overhead
Getting rid of the communications overhead The core feature of BXI is a full hardware-encoded communication management system, which enables CPUs to be fully dedicated to computational tasks while communications are independently managed by BXI.
As a result, contrary to other commonly used networks, BXI can deliver high communication throughput even when the system is under heavy computation stress.
BXI hardware primitives map directly to communication libraries such as MPI (Message Passing Interface) and PGAS (Partitioned Global Address Space). Thanks to this hardware acceleration, BXI delivers the highest level of communication performance for HPC applications, at full scale, characterized by high bandwidth, low latency and high message rates.
The BXI architecture is based on the Portals 4 communication library. This enables full optimization for all MPI communication types, including the latest MPI-2 and MPI-3 extensions and PGAS. The Portals 4 non-connected protocol guarantees a minimum constant memory footprint, irrespective of system size.
Quality of service
BXI quality of service (QoS) enables the definition of several virtual networks and ensures, for example, that bulky I/O messages do not impede small data message flow. In addition, BXI adaptive routing capabilities dynamically avoid communication bottlenecks.
Reliability and resilience
For high reliability, BXI implements both end-to-end and link-level error checking and retransmission. Furthermore, all ASIC parts feature ECC schemes for error detection and correction. These mechanisms ensure continuity of service in case of a transient or permanent failure (on link or switch).
The BXI fabric relies on two types of ASICs as its building blocks, a Network Interface Controller (NIC) and a switch, and comes with its complete software suite. BXI switches are managed through a distributed and out-of-band fabric management suite allowing to scale up to 64K nodes. Out-of-band management eliminates any interference of the management traffic with the applications traffic.