Posibilidad de añadir alguna información al Informe de alta

Several works focus on deriving an upperbound of WCTT adapting network calculus [101] to model wNoCs [81]. Network calculus relies on the determination of arrival curves of the applications running in the system to determine an actual upperbound of WCTT. While these

Chapter 4. Modeling High-Performance wNoCs for CRTES 75 approaches allow providing tight WCTT estimates, as WCTT is adapted to the exact network load conditions, using per-application arrival curves reduces time composability, since WCTT estimates depend on the load corrunners put on the NoC. Another set of works focuses on determining wNoC packets WCTT by considering worst-case conditions, first with assuming limitations on the packet-injection rate [80]. However, the bounds provided in [80] required assuming packet injection is limited. In later works [3,102] this limitation is removed.

In this line of work, Dasari et. al [82] achieve tighter WCTT bounds – than those derived with [3, 102] – based on the following two observations: (1) flows injection rate is inherently limited by the speed at which the processor pipeline can process request-generating instructions (e.g load or stores); and (2) packets of a given flow do not always contend with the flow under analysis due to the existing release time of their request of a flow/core. Regarding observation (1) we have shown that while limiting the injection effectively reduces WCTT values the actual contention that packets in the NoC suffer remains unaltered, hence producing no effect on WCD. Regarding observation (2) knowing what is the actual interval between consecutive requests in every flow in the network breaks the time composability requirement.

4.9

Conclusions

This chapter analyzes the suitability of applying wNoCs in the real-time embedded domain. To do so, we propose a new metric to account for the impact that NoC interferences coming from different requesters have on the WCET estimates: the Worst-Contention Delay (WCD), which replaces the traditional metric, the Worst-Case Traversal Time (WCTT). Moreover, we derive an analytical model that computes time-composable WCD bounds (W CD) based on common wNoC design parameters including flits-per-packet, number of virtual channels and queue size in the router. W CD is computed based on a wNoC parameter taxonomy that identifies those parameters that must be fixed in order to provide trustworthy and composable WCD bounds; and those allowing certain flexibility (objectives O1 and O3).

Our W CD model allows evaluating a wide range of existing COTS high-performance wNoCs. To that end, we apply the model considering the design parameters of two wNoCs deployed in real processors: the Tilera-Gx36 and the 48-core Intel SCC (objective O1). Our analysis shows that considering WCD rather than WCTT reduces WCET estimates by around 2.5x for Tilera and ISCC on average (objective O3).

Chapter 5

Improving Performance

Guarantees in Wormhole Mesh

NoC Designs

Wormhole-based Network on Chips (wNoCs) are deployed in high-performance many-core pro- cessors due to their physical scalability and low-cost. Delivering tight and time composable Worst-Case Execution Time (WCET) estimates for applications as needed in safety-critical real-time embedded systems is challenged by wNoCs due to their distributed nature. In this chapter we propose a bandwidth control mechanism for wNoCs that enables the computation of tight time-composable WCET estimates with low average performance degradation and high scalability. Our evaluation with the EEMBC automotive suite and an industrial real-time parallel avionics application confirms so.

5.1

Introduction

Critical Real-Time Embedded Systems (CRTES) industry is gradually shifting towards multi- and manycore processors to satisfy the performance needs of complex safety-related functions. This transition challenges the derivation of time-composable WCET estimates, i.e. tasks’ execution time bounds that are independent of the load that co-running tasks put on shared resources. Time-composable WCET estimates enable incremental verification [103] by allowing each system component to be subject to formal timing verification in isolation and independently from other components.

From an end-user perspective, the deployment of manycores in CRTES, as stated in Section 1.3 requires following properties:

• UserReq1 : Manycores should facilitate deriving tight WCET estimate so that high (guar- anteed) performance is provided (objectives O1 and O3);

Chapter 5. Improving Performance Guarantees in Wormhole Mesh NoC Designs 78 • UserReq2 : Manycores must facilitate deriving time composable WCET estimates (objective

O1);

• UserReq3 : Manycores should also provide high average performance for some applications (objective O1);

• UserReq4 : Manycores for real-time should use technology as close as possible to Commercial- Off-The-Shelf (COTS) (high-performance) technology to ease their adoption (objective O1). The low manycore demand of safety-critical real-time systems, w.r.t. the mainstream market, calls for reducing the need for customized real-time technology.

This chapter tackles the fulfillment of the above requirements on Network on Chip (NoC) designs, as it is one of the manycore shared resources with the highest impact on average performance and WCET. Concretely, we consider wNoC mesh as a candidate NoC solution as it is widely accepted in the high-performance market due to its physical scalability and low cost [18][60]. The high-performance requirements (UserReq3 ) are already fulfilled by wNoCs as they are designed for high-performance systems. UserReq2 for real-time applications requires time-composable Worst-Case Traversal Time (WCTT), i.e. WCTT not affected by the load contender tasks put on the wNoC. Typically, latency bounds for wNoCs are reffered as WCTT. wNoCs can also meet this by using time-analyzable arbitration policies [86][85] and applying the model in [3]. This chapter makes the following contribution:

• We show that current wNoCs fail to achieve tight WCTT (UserReq1 ), which negates their benefits. In particular we show that (i) WCTT values derived for current wNoCs poorly scale with network size – even for small networks; and (ii) the WCTT derived for a task depends on the maximum allowed packet size and poorly scales with it. Further, current wNoCs do not necessarily impose a limit on the packet size and leave that to the protocol on top of the network (e.g. AMBA [88]).

• We propose a new time-composable wNoC design relying on concepts developed for high- performance wNoCs, hence achieving UserReq4 and objective O1. Our design focuses on controlling the network bandwidth (the main factor affecting WCTT ) to provide a fair guaranteed bandwidth distribution across the different communication flows. Bandwidth control is exercised at two levels. At local level, we ensure fairness by providing a WCTT- aware Packetization (WaP) that makes real-time guarantees independent of contenders packet size. At global level, we provide fairness across contenders by performing a WCTT- aware Weighted (WaW) round-robin arbitration.

• We evaluate WaW+WaP on a 64-core manycore architecture with cores accessing memory controllers through a wNoC. We use EEMBC [43] autobench and an avionics real-time parallel application provided by Honeywell [30] (objective O4). We show that our design significantly decreases WCET estimates for the parallel application by a factor of 4.8_{× to} 9.5_{× depending on the number of flits per packet. For single-threaded applications WCET} decreases by 230× on average across all cores and by 1.4× w.r.t 25% of the best cores of the baseline NoC.

Chapter 5. Improving Performance Guarantees in Wormhole Mesh NoC Designs 79 Note that proposals made in this chapter can be orthogonally applied to Worst-Contention Delay (WCD) metric proposed in Chapter 4. We opt for comparison to the WCTT and state- of-the art wNoC proposals[58], as the main benefit of the proposed WaW+WaP is the fairness and it is affecting both approaches in the same manner, but comparison using WCTT better highlights the effectiveness of the solution.