The BXI routing architecture for exascale supercomputer

Vigneras, Pierre; Quintin, Jean-Noël

doi:10.1007/s11227-016-1755-2

Cited by 8 publications

(6 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, static reconfiguration ensures deadlock freedom as there is no moment that two different routing functions, the R old and R new , coexist. As the time it takes to reconfigure the network can range in the order of minutes, especially in large subnets as we and others show in Paper V and [40,41], dynamic reconfiguration is generally preferred over the static. In dynamic reconfiguration the traffic keeps flowing while the network gets reconfigured, but the reconfiguration must be done carefully in order to ensure the deadlock freedom of the network during the transition from R old to R new when the new routes are distributed to all switches.…”

Section: Network Reconfigurationmentioning

confidence: 90%

See 1 more Smart Citation

A Novel Query Caching Scheme for Dynamic InfiniBand Subnets

Tasoulas

Gran

Johnsen³

et al. 2015

2015 15th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing

View full text Add to dashboard Cite

The advent of the Internet of Things, sensor and social networks, to mention just a few examples, all contribute towards the solid establishment of the Big Data era. High Performance Computing (HPC) becomes necessary for the efficient processing of the massive amounts of data our society generates, and cloud computing is a critical component to deliver this processing power to a broader audience that cannot afford to acquire and maintain such complex computing systems themselves. However, HPC specific technology and performance is not yet apt to be delivered efficiently over highly flexible and dynamic environments, as typically are the virtualized cloud infrastructures.In this thesis, we address challenges that arise in high performance dynamic cloud environments, that are equipped with HPC specific technology, in the context of networking and virtualization. We use InfiniBand, a high performance lossless interconnection network as the basis of our research, and first show that lossless networks pose prime challenges when the nature of the infrastructure is very dynamic, i.e. exhibits continuous changes. Then we propose a network I/O virtualization architecture, the InfiniBand vSwitch architecture, that can make lossless network technologies more favorable in the cloud. Moreover, we propose different network reconfiguration methods to enable performance-driven reconfigurations in very large network topologies that are commonly found in data centers. Performance-driven reconfigurations are frequently needed to adapt to unpredictable workload changes resulting from the shared and on-demand nature of a cloud platform, or when cloud providers employ live migration of virtual machines to optimize the resource usage of their infrastructure. Last but not least, we propose a new Quality-of-Service metric, called delay, to capture the directly observable service degradation in consolidated cloud environments. We suggest that the delay can be used as a direct service level agreement metric between cloud providers and cloud tenants. AcknowledgementsFirst, I would like to express my gratitude to my three supervisors; Ernst Gunnar Gran, Tor Skeie and Kyrre Begnum. Without their positiveness, supervision, suggestions and encouragement, my PhD journey would not have been so fruitful and enjoyable during both the good and hard times. Then I would like to thank Bjørn Dag Johnsen from Oracle Norway, our closest collaborate in the ERAC project, the project that mainly funded this thesis, for his enthusiastic attitude and participation in the important discussions that shaped the direction of my work and introduced industrial relevance. My appreciation goes as well to Feroz Zahid for being the most easygoing colleague whom I could imagine sharing office with, and a brilliant and ambitious associate, and Sven-Arne Reinemo for his supervision during my early days.A special thank you must be directed to Hårek Haugerud, Anis Yazidi, Hugo Lewi Hammer, Laurence Marie Anna Habib and the whole NETSYS group at the Oslo and Akershus...

show abstract

Section: Network Reconfigurationmentioning

confidence: 90%

“…This task can take several minutes depending on the network size and topology [40,41]. Normally, HPC clusters are static environments and once deployed the SM rarely intervenes to reconfigure the network unless critical faults are detected.…”

Section: Research Challenges Addressed In This Thesismentioning

confidence: 99%

A Novel Query Caching Scheme for Dynamic InfiniBand Subnets

Tasoulas

Gran

Johnsen³

et al. 2015

2015 15th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing

View full text Add to dashboard Cite

show abstract

“…For rack-scale interconnects, a custom protocol can be used to achieve this distribution of a common clock. For instance, the BXI interconnect from Atos broadcasts a common clock with an accuracy of 2 µsecs [22]. As said above, broadcasting the RRP clock event is straightforward in NOCs.…”

Section: Discussionmentioning

confidence: 99%

Accurate Congestion Control for RDMA Transfers

Giannopoulos

Chrysos

Mageiropoulos

et al. 2018

2018 Twelfth IEEE/ACM International Symposium on Networks-on-Chip (NOCS)

View full text Add to dashboard Cite

High-performance interconnects need congestion control to deal with traffic bursts. In this paper, we propose ACCurate, a congestion control protocol that assigns exact maxmin fair rates to flows, without relying on costly per-flow state inside the network. ACCurate keeps the backlogs outside of the network, protects innocent flows, and promptly recovers the flows' rates after congestive episodes. Comparisons with TCP and PAUSE-only RDMA under datacenter-resembling workloads further show that ACCurate provides up to 10x faster flow completion times. ACCurate relies on simple hardware that can be readily implemented inside switches. In our implementation, the additional circuitry needed in a 16×16 switch occupies less than 2% of FPGA resources.

show abstract

“…More recently, the Bull EXascale Interconnect (BXI) [10] has followed a similar approach. They use a 2-stage routing strategy [21]: first an off-line algorithm calculates the paths between each source and destination. These paths are deterministic and populated into the routing tables during system start-up (could be done arithmetically).…”

Section: Related Workmentioning

confidence: 99%

“…buffers, crossbar, flow control mechanisms, etc). This is because tens of thousands of entries need to be stored to be able to reach all the nodes of the system [21]. In addition, routing tables create other scalability issues.…”

Section: Introductionmentioning

confidence: 99%

A CAM-Free Exascalable HPC Router for Low-Energy Communications

Concatto

Pascual

Navaridas

et al. 2018

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Power consumption is the main hurdle in the race for designing Exascale-capable computing systems which would require deploying millions of computing elements. While this problem is being addressed by designing increasingly more power-efficient processing subsystems, little effort has been put on reducing the power consumption of the interconnection network. This is precisely the objective of this work, in which we study the benefits, in terms of both area and power, of avoiding costly and power-hungry CAM-based routing tables deep-rooted in all current networking technologies. We present our custom-made, FPGAbased router based on a simple, arithmetic routing engine which is shown to be much more power-and area-efficient than even a relatively small 2K-entry routing table which requires as much area and one order of magnitude more power than our router.

show abstract

The BXI routing architecture for exascale supercomputer

Cited by 8 publications

References 30 publications

A Novel Query Caching Scheme for Dynamic InfiniBand Subnets

A Novel Query Caching Scheme for Dynamic InfiniBand Subnets

Accurate Congestion Control for RDMA Transfers

A CAM-Free Exascalable HPC Router for Low-Energy Communications

Contact Info

Product

Resources

About