NetFC: Enabling Accurate Floating-point Arithmetic on Programmable Switches

Cui, Penglai; Pan, Heng; Li, Zhenyu; Wu, Jiaoren; Zhang, Shengzhuo; Yang, Xingwu; Guan, Hongtao; Xie, Gaogang

doi:10.1109/icnp52444.2021.9651946

Cited by 15 publications

(6 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The worker deems the gradient packet has been lost, so it retransmits the packet. Without a local 3 While this on-chip memory demand is acceptable in practice, we can also leverage the prefix-based compression proposed in [15] to further reduce the memory consumption. recording in the switch, the switch will wrongly aggregate the gradients twice.…”

Section: System Reliabilitymentioning

confidence: 99%

See 1 more Smart Citation

Libra: In-network Gradient Aggregation for Speeding up Distributed Sparse Deep Training

Pan¹,

Cui²,

li³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Distributed sparse deep learning has been widely used in many internet-scale applications. Network communication is one of the major hurdles for the training performance. Innetwork gradient aggregation on programmable switches is a promising solution to speed up the performance. Nevertheless, existing in-network aggregation solutions are designed for the distributed dense deep training, and fall short when used for the sparse deep training. To address this gap, we present Libra based on our key observation of the extremely biased update frequency of parameters in distributed deep sparse training. Specifically, Libra offloads only the aggregation for "hot" parameters that are updated frequently onto programmable switches. To enable this offloading and achieve high aggregation throughput, we propose solutions to address the challenges related to hot parameter identification, parameter orchestration, floating-point summation on switches as well as system reliability. We implemented Libra on Intel Tofino switches and integrated it with PS-lite. Finally, we evaluate Libra's performance through extensive experiments and show that Libra can speed up the gradient aggregation by 1.5∼4×. miTable Find σ(z)=log2(1+2z) z=log2(y)-log2(x) t=log2(x)+s(z)

show abstract

Section: System Reliabilitymentioning

confidence: 99%

“…Rather than using the aforementioned floating-to-integer method that may introduce accuracy loss, we propose a table-lookup method that enables the on-the-fly floating-point summation for 32-bit floats. Our method is inspired by [15] that implements floatingpint operations for 16-bit floats.…”

mentioning

confidence: 99%

Libra: In-network Gradient Aggregation for Speeding up Distributed Sparse Deep Training

Pan¹,

Cui²,

li³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Most of the programmable data plane architecture [63] can not support complex mathematical operations such as multiplication and division. The solution is to adopt the approximation techniques [64] of bit-shifting with adders and table lookups. In addition, more complex operations such as exponential and logarithmic can also be realized [45] based on the approximation techniques with binomial series expansion.…”

Section: B P4mentioning

confidence: 99%

Tabular Interpolation Approach Based on Stable Random Projection for Estimating Empirical Entropy of High-Speed Network Traffic

et al. 2022

View full text Add to dashboard Cite

The empirical entropy of the network flow attributes is an essential measure for identifying anomalous network traffic. However, computing the exact entropy values for high-speed networks in realtime is computationally expensive. Accordingly, the present study replaces the complex computations of existing stable random projection methods for entropy estimation with a simple table lookup procedure. Notably, the size of the lookup table is reduced through a piece-wise linear interpolation heuristic in order to facilitate the implementation of the proposed scheme in resource-constrained pipeline environments. The proposed architecture enables entropy estimation to be performed using both the Log-Mean Estimator (LME) method and the New Estimator of Compressed Counting (NECC) algorithm reported in the literature. The feasibility of the proposed approach is verified empirically using both real-world network traffic traces and synthetic data streams. Moreover, the practical applicability is demonstrated via stream-based implementation in the programmable data planes of the NetFPGA-Plus framework and a Tofino P4 switch, respectively. The results indicate that the proposed tabulation-based entropy estimation scheme allows minimum-sized Ethernet frames to be processed with a wire speed of up to several hundred gigabits per second.

show abstract

“…Recently, new methods for computing elementary operations log and exp have been proposed in [30]. We can also mention [31] that implements floating point arithmetic in-network with 99.94% accuracy in the worst case. N2Net [32] and BaNANA Split [33] have shown implementations of binary neural networks on the dataplane.…”

Section: B In-network Real Value Computationmentioning

confidence: 99%

NetREC: Network-wide in-network REal-value Computation

Jose

Lazri

Jérôme

et al. 2022

2022 IEEE 8th International Conference on Network Softwarization (NetSoft)

View full text Add to dashboard Cite

The current generation of networks empowers the use of programmable switches whose behaviour can be defined using languages like P4. Nevertheless, these languages do not support network-wide deployment of stateful real-value functions. This paper presents NetREC, an extension of RMT programmable data planes designed to enable stateful real-value functions computation across multiple switches.NetREC first decomposes the real-value functions into a dependency graph of elementary operations that are distributed among the network. This distribution is carried out by dynamically generating and solving an integer linear program. We deploy a prototype of NetREC on a network of Tofino switches and demonstrate its capability of computing recursive real-value functions like exponential weighted moving average.

show abstract

NetFC: Enabling Accurate Floating-point Arithmetic on Programmable Switches

Cited by 15 publications

References 20 publications

Libra: In-network Gradient Aggregation for Speeding up Distributed Sparse Deep Training

Libra: In-network Gradient Aggregation for Speeding up Distributed Sparse Deep Training

Tabular Interpolation Approach Based on Stable Random Projection for Estimating Empirical Entropy of High-Speed Network Traffic

NetREC: Network-wide in-network REal-value Computation

Contact Info

Product

Resources

About