Sujet : ICCD 2017 notification for paper 347 De : ICCD 2017 Date : 31/08/2017 17:46 Pour : Frank Singhoff Dear Frank Singhoff, The ICCD 2017 program committee is delighted to inform you that your paper 347, entitled "DAS: an efficient NoC Router for Mixed-Criticality Real-Time Systems", has been accepted to appear at the conference as a short-length paper (4 pages maximum) and a shorter-duration oral presentation. Your paper was one of 35 accepted short papers in addition to 75 accepted full length papers out of 258 submissions. Congratulations! Reviews on your paper are appended to this email. Instructions on preparing the camera ready version of your paper will be communicated shortly. Please note that each accepted paper MUST have a unique full registration and an author present a poster at the conference. This is a mandatory requirement for your paper to be included in the official proceedings. Instructions on registering for the conference will be communicated shortly. Thank you for your submission. We will see you at ICCD 2017 in Boston, MA, USA in November. Best Regards, Ozgur Sinanoglu and Umit Ogras ICCD 2017 Program Chairs ----------------------- REVIEW 1 --------------------- PAPER: 347 TITLE: DAS: an efficient NoC Router for Mixed-Criticality Real-Time Systems AUTHORS: Mourad Dridi, Stéphane Rubini, Mounir Lallali, Johanna Sepulveda, Frank Singhoff and Jean-Philippe Diguet Overall evaluation: 2 (accept) ----------- Overall evaluation ----------- This paper presents a new virtual channel router for mixed criticality systems. Based on the criticality flow, it uses wormhole and store and forward techniques. The results are very good and accurate. The architecture could be explained in more detail. ----------------------- REVIEW 2 --------------------- PAPER: 347 TITLE: DAS: an efficient NoC Router for Mixed-Criticality Real-Time Systems AUTHORS: Mourad Dridi, Stéphane Rubini, Mounir Lallali, Johanna Sepulveda, Frank Singhoff and Jean-Philippe Diguet Overall evaluation: -1 (weak reject) ----------- Overall evaluation ----------- Pros: + good writing with strong motivation. + clear explanation of using SAF for high-critical flows, and wormhole for low-critical flows. + high-critical flow is only preempted at the packet level so it is able to schedule of flows for improving the worst case communication time. + high-critical flows can preempt low-critical flows at the flit level so that they can meet their deadlines. Cons: - with N VCs per router, a physical link is allowed to be shared by maximum of N high-critical flows. - not clear how the proposal is scaled to support more than two criticality levels. - multiple low-critical flows sharing the same only one VC can cause protocol deadlock. - preempted by high-critical flows can cause starvation for low-critical flows This paper presents DAS - a NoC router design for mixed-criticality real-time systems. DAS router has N+1 VCs in which N VCs are dedicated to N high-critical flows, and one VC is shared for all low-critical flows. High-critical flows use store and forward (SAF) switching technique while low-critical flows use wormhole technique. High-critical flows can preempt low-critical flows at the flit levels, while high-critical flows themselves have the same priority and are only preempted at the packet level. There are 3 modes per router I/O port which are depending on the arrival of flows at the port. The proposed router is shown to reduced 80% latency of high-critical flows compared to VC routers while only has 2.5% area overhead. Main contributions: . High-critical flows can preempt low-critical flows at the flit levels to be able to meet their deadlines. . Among high-critical flows, preemption is at the packet level which allow the users to schedule those flows for improving their worst case communication time. . Using SAF switching technique for high-critical flows makes sense because high-critical flows (such as sensor data and control signals) are normally transferred in small size packets. Concerns: . How can the proposed technique be scaled to support more than two criticality levels? For example, if the system has 3 criticality levels of flows: high, medium and low; how are the VCs organized and what is the preemption scheme for these flows? . Because DAS router only supports N VCs for N high critical flows, it requires a special mapping algorithm which ensures that less than N critical flows share a given physical link. That mapping algorithm is a NP-Hard problem and doesn’t always guarantees to work for all applications which different communication characteristics. What will happen if the algorithm leaves a few physical links with more than N high critical flows? Does DAS still meet real-time deadlines for those flows? If not, what is the solution to deal with this issue? . How does the router identify the criticality level of an incoming flow? Do we need to add the flow identification bits to each packet? . Multiple low-critical flows share only one VC per router port; this can cause protocol deadlock due to a loop among these low-critical flows. . Because high-critical flows always preempt low-critical flows, that may cause starvation issue for low-critical flows. In a worse case, if a physical link always have high-critical flows coming from different sources, the low-critical flows on that link could never move. How do the authors address this issue? . Router area is also decided by target clock frequency; so what is the clock frequency the authors used to synthesize the routers? In summary, the proposed idea allows reducing the latency of high-critical flows while scarifying some bandwidth for low-critical flows. It needs a special mapping algorithm to work. The authors also need to address the starvation issue for low-critical flows. ----------------------- REVIEW 3 --------------------- PAPER: 347 TITLE: DAS: an efficient NoC Router for Mixed-Criticality Real-Time Systems AUTHORS: Mourad Dridi, Stéphane Rubini, Mounir Lallali, Johanna Sepulveda, Frank Singhoff and Jean-Philippe Diguet Overall evaluation: 1 (weak accept) ----------- Overall evaluation ----------- Interesting work. However, no correctness argument is given. How do the authors guarantee the absence of deadlock? Experiments are presented for only 15% network use rate. What happens for higher numbers? Also, the network traffic is generated randomly for the experiments, as opposed to taken from actual applications, and this raises another concern. ----------------------- REVIEW 4 --------------------- PAPER: 347 TITLE: DAS: an efficient NoC Router for Mixed-Criticality Real-Time Systems AUTHORS: Mourad Dridi, Stéphane Rubini, Mounir Lallali, Johanna Sepulveda, Frank Singhoff and Jean-Philippe Diguet Overall evaluation: -1 (weak reject) ----------- Overall evaluation ----------- This paper proposes a routing technique for mixed-criticality real-time systems, which uses wormhole for low critical flow, store and forward for high critical flow. Overall the idea is clearly described. However, currently there are a few drawbacks. 1. The claim "none of them (earlier NoCs) can handle both high and low-critical flows" is not true. Lots of dynamic/prioritized/reconfigurable NoCs can handle flows with different timing constrains. They may not emphasize “mixed-criticality”, but I do not see why none of them can handle mixed-criticality flows. 2. It is true that WCCT may be hard to measure, and it is hard to define N for arbitrary on-chip traffic. But the assumption “at most N high-critical flows sharing the same link” over simplifies the problem. This assumption make the calculation of WCCT much easier, but it is hard to believe this would hold in real cases. Such as in the simulation setup, only 3 VC for a 4*4 mesh, which means at most only 2 high-critical flows can share the same link. I feel this is not very representative. 3. Another assumption that high critical flow always has small packet, while low critical flow always has larger ones, also favor the proposed scheme. It means applying SAF for small packet and wormhole for large packet. In addition, the simulation set up is over simplified, only one high-critical flow is assigned. These make the result less appealing.