The growing adoption of hardware accelerators driven by their intelligent compiler and runtime system counterparts has democratized ML services and precipitously reduced their execution times. This motivates us to shift our attention to efficiently serve these ML services under distributed settings and characterize the overheads imposed by the RPC mechanism ('RPC tax') when serving them on accelerators. The RPC implementations designed over the years implicitly assume the host CPU services the requests, and we focus on expanding such works towards accelerator-based services. While recent proposals calling for SmartNICs to take on this task are reasonable for simple kernels, serving complex ML models requires a more nuanced view to optimize both the data-path and the control/orchestration of these accelerators. We program today's commodity network interface cards (NICs) to split the control and data paths for effective transfer of control while efficiently transferring the payload to the accelerator. As opposed to unified approaches that bundle these paths together, limiting the flexibility in each of these paths, we design and implement SplitRPC - a control + data path optimizing RPC mechanism for ML inference serving. SplitRPC allows us to optimize the datapath to the accelerator while simultaneously allowing the CPU to maintain full orchestration capabilities. We implement SplitRPC on both commodity NICs and SmartNICs and demonstrate how GPU-based ML services running different compiler/runtime systems can benefit. For a variety of ML models served using different inference runtimes, we demonstrate that SplitRPC is effective in minimizing the RPC tax while providing significant gains in throughput and latency over existing kernel by-pass approaches, without requiring expensive SmartNIC devices.
more »
« less
SplitRPC: A {Control + Data} Path Splitting RPC Stack for ML Inference Serving
The growing adoption of hardware accelerators driven by their intelligent compiler and runtime system
counterparts has democratized ML services and precipitously reduced their execution times. This motivates
us to shift our attention to efficiently serve these ML services under distributed settings and characterize
the overheads imposed by the RPC mechanism (‘RPC tax’) when serving them on accelerators. The RPC
implementations designed over the years implicitly assume the host CPU services the requests, and we focus
on expanding such works towards accelerator-based services. While recent proposals calling for SmartNICs
to take on this task are reasonable for simple kernels, serving complex ML models requires a more nuanced
view to optimize both the data-path and the control/orchestration of these accelerators. We program today’s
commodity network interface cards (NICs) to split the control and data paths for effective transfer of control
while efficiently transferring the payload to the accelerator. As opposed to unified approaches that bundle
these paths together, limiting the flexibility in each of these paths, we design and implement SplitRPC - a
{control + data} path optimizing RPC mechanism for ML inference serving. SplitRPC allows us to optimize the
datapath to the accelerator while simultaneously allowing the CPU to maintain full orchestration capabilities.
We implement SplitRPC on both commodity NICs and SmartNICs and demonstrate how GPU-based ML services
running different compiler/runtime systems can benefit. For a variety of ML models served using different
inference runtimes, we demonstrate that SplitRPC is effective in minimizing the RPC tax while providing
significant gains in throughput and latency over existing kernel by-pass approaches, without requiring
expensive SmartNIC devices.
more »
« less
- Award ID(s):
- 1763681
- NSF-PAR ID:
- 10450471
- Date Published:
- Journal Name:
- Proceedings of the ACM on measurement and analysis of computing systems
- ISSN:
- 2476-1249
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Model-serving systems expose machine learning (ML) models to applications programmatically via a high-level API. Cloud plat- forms use these systems to mask the complexities of optimally managing resources and servicing inference requests across multi- ple applications. Model serving at the edge is now also becoming increasingly important to support inference workloads with tight latency requirements. However, edge model serving differs substan- tially from cloud model serving in its latency, energy, and accuracy constraints: these systems must support multiple applications with widely different latency and accuracy requirements on embedded edge accelerators with limited computational and energy resources. To address the problem, this paper presents Dělen,1 a flexible and adaptive model-serving system for multi-tenant edge AI. Dělen exposes a high-level API that enables individual edge applications to specify a bound at runtime on the latency, accuracy, or energy of their inference requests. We efficiently implement Dělen using conditional execution in multi-exit deep neural networks (DNNs), which enables granular control over inference requests, and evalu- ate it on a resource-constrained Jetson Nano edge accelerator. We evaluate Dělen flexibility by implementing state-of-the-art adapta- tion policies using Dělen’s API, and evaluate its adaptability under different workload dynamics and goals when running single and multiple applications.more » « less
-
FlexTOE is a flexible, yet high-performance TCP offload engine (TOE) to SmartNICs. FlexTOE eliminates almost all host data-path TCP processing and is fully customizable. FlexTOE interoperates well with other TCP stacks, is robust under adverse network conditions, and supports POSIX sockets. FlexTOE focuses on data-path offload of established connections, avoiding complex control logic and packet buffering in the NIC. FlexTOE leverages fine-grained parallelization of the TCP data-path and segment reordering for high performance on wimpy SmartNIC architectures, while remaining flexible via a modular design. We compare FlexTOE on an Agilio-CX40 to host TCP stacks Linux and TAS, and to the Chelsio Terminator TOE. We find that Memcached scales up to 38% better on FlexTOE versus TAS, while saving up to 81% host CPU cycles versus Chelsio. FlexTOE provides competitive performance for RPCs, even with wimpy SmartNICs. FlexTOE cuts 99.99th-percentile RPC RTT by 3.2× and 50% versus Chelsio and TAS, respectively. FlexTOE's data-path parallelism generalizes across hardware architectures, improving single connection RPC throughput up to 2.4× on x86 and 4× on BlueField. FlexTOE supports C and XDP programs written in eBPF. It allows us to implement popular data center transport features, such as TCP tracing, packet filtering and capture, VLAN stripping, flow classification, firewalling, and connection splicing.more » « less
-
Dense matrix multiply (MM) serves as one of the most heavily used kernels in deep learning applications. To cope with the high computation demands of these applications, heterogeneous architectures featuring both FPGA and dedicated ASIC accelerators have emerged as promising platforms. For example, the AMD/Xilinx Versal ACAP architecture combines general-purpose CPU cores and programmable logic (PL) with AI Engine processors (AIE) optimized for AI/ML. An array of 400 AI Engine processors executing at 1 GHz can theoretically provide up to 6.4 TFLOPs performance for 32-bit floating-point (fp32) data. However, machine learning models often contain both large and small MM operations. While large MM operations can be parallelized efficiently across many cores, small MM operations typically cannot. In our investigation, we observe that executing some small MM layers from the BERT natural language processing model on a large, monolithic MM accelerator in Versal ACAP achieved less than 5% of the theoretical peak performance. Therefore, one key question arises: How can we design accelerators to fully use the abundant computation resources under limited communication bandwidth for end-to-end applications with multiple MM layers of diverse sizes? We identify the biggest system throughput bottleneck resulting from the mismatch of massive computation resources of one monolithic accelerator and the various MM layers of small sizes in the application. To resolve this problem, we propose the CHARM framework to compose multiple diverse MM accelerator architectures working concurrently towards different layers within one application. CHARM includes analytical models which guide design space exploration to determine accelerator partitions and layer scheduling. To facilitate the system designs, CHARM automatically generates code, enabling thorough onboard design verification. We deploy the CHARM framework for four different deep learning applications, including BERT, ViT, NCF, MLP, on the AMD/Xilinx Versal ACAP VCK190 evaluation board. Our experiments show that we achieve 1.46 TFLOPs, 1.61 TFLOPs, 1.74 TFLOPs, and 2.94 TFLOPs inference throughput for BERT, ViT, NCF, MLP, respectively, which obtain 5.40x, 32.51x, 1.00x and 1.00x throughput gains compared to one monolithic accelerator.more » « less
-
As customized accelerator design has become increasingly popular to keep up with the demand for high performance computing, it poses challenges for modern simulator design to adapt to such a large variety of accelerators. Existing simulators tend to two extremes: low-level and general approaches, such as RTL simulation, that can model any hardware but require substantial effort and long execution times; and higher-level application-specific models that can be much faster and easier to use but require one-off engineering effort.This work proposes a compiler-driven simulation workflow that can model configurable hardware accelerator. The key idea is to separate structure representation from simulation by developing an intermediate language that can flexibly represent a wide variety of hardware constructs. We design the Event Queue (EQueue) dialect of MLIR, a dialect that can model arbitrary hardware accelerators with explicit data movement and distributed event-based control; we also implement a generic simulation engine to model EQueue programs with hybrid MLIR dialects representing different abstraction levels. We demonstrate two case studies of EQueue-implemented accelerators: the systolic array of convolution and SIMD processors in a modern FPGA. In the former we show EQueue simulation is as accurate as a state-of-the-art simulator, while offering higher extensibility and lower iteration cost via compiler passes. In the latter we demonstrate our simulation flow can guide designer efficiently improve their design using visualizable simulation outputs.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
