network performance analysis and management for

R0(t) is the total amount of computation delivered up to time, with the associated computation delivered Rb(t).CandC0 are the capacity function and residual capacity functions that describe the total processing capacity under full load and the residual processing capacity , respectively .CandC0 are bounded by the delivery curve β and the residual distribution curve β0, reprinted from [50]. Application communications are indicated by solid black arrows, and out-of-band communications are indicated by red dotted arrows.

Motivation

In any distributed CPS, the system's network performance is affected by the system's physical environment as it affects the network. However, incorporating the physical dynamics into the model of the system's network resources only solves half the problem.

Challenges

Design-Time Network Performance Analysis of Dis-

These routes may be defined at design time and remain constant for the duration of the system, or may be unknown at design time, changing dynamically during the runtime of the system. Dynamic routing is difficult to analyze for accurate performance prediction because the routes used by traffic may be unknown at design time.

Run-Time Network Resource Monitoring and Manage-

For resource-constrained systems, no processor or memory resources should be wasted, but without accurate and precise design-time analysis, systems should conservatively overestimate network resource requirements. For application/system data flows in the network that require strict and/or real-time guarantees on temporal properties, design-time analysis is critical.

Organization

We define systems to be periodic if the system's data production rate (or consumption rate) is a periodic function of time. Consider the scenario that the remaining capacity of the system is less than the size of the buffer, i.e. R[Tp]

Related Work

Part 1: Design-Time Network Analysis and Performance Predic-

Performance Analysis Through Network Simulation/Em-
Analytical Approaches to Network Analysis

For example, there is a temporal disconnect between arrival/service curves and actual application or system performance. The authors in [4] also incorporate more accurate network models to derive tighter performance bounds using grid computing.

Figure 1: Example wide sense increasing functions, reprinted from [34].

Part 2: Run-Time Network Monitoring and Management

Infrastructural Approaches for Network Management . 27

Two of the main infrastructure methods for system network service management are DiffServ[41][20] and Intserv[5]. FQMM focuses on enabling accurate node resource provisioning and enabling node mobility by dynamically reallocating the roles of each of the nodes in the network. Resource provisioning for network flows borrows ideas from both IntServ and DiffServ by combining IntServ's per-flow fragmentation for high-priority flows, while lower-priority flows are provisioned on a class-by-class basis, as in DiffServ.

For such networks, the physical dynamics of the nodes in the cluster are well understood and predictable, and therefore the network dynamics can also be reasonably predictable. Integrating the physical dynamics of the network into the modeling and analysis tools improves the performance of the systems without affecting their reliability.

Precise Modeling of Application Network Traffic and System

Problem
Mathematical Formalism
Accuracy and Precision
Assumptions Involved
Factors Impacting Analysis

To model system network capability and application traffic patterns, we developed a network modeling paradigm similar to Network Calculus traffic arrival curves and traffic shaper service curves. For these experiments, we configured the traffic shaping node to control the data rate of application data on the network interface according to the network profile provided by the system. As can be seen in the table, the predictions made by our analysis techniques are tight and conservative bounds on the actual performance of the application on the experimental system.

Such concepts are especially important with respect to how they influence the behavior of the network software stack, including user-space applications. However, as with any system, there is a trade-off between model accuracy and analysis time and results.

Figure 9: Representative example demonstrating convolution of the data vs. time profiles that the application requires and that the system provides

Analysis of Periodic Systems

Proving the Minimum Analysis for System Stability

First, we see that in system (1) the predicted required buffer size does not change regardless of the number of periods over which we analyze the system. We can determine the hyperperiod of the system as thelcmof input function period and the service function period,Tp=lcm(TS,TI). Therefore, the amount of data in the buffer at the end of the first period, t =Tp, is the amount of data that entered the system on input function I but was unable to be serviced by S.

Data in the buffer at the beginning of the period can be compared to the system's remaining capacity R, since the system's remaining capacity indicates how much additional data it can send in that period. This relationship means that if the remaining capacity of the system that exists after all the period's required traffic has been handled is equal to or greater than the size of the buffer at the end of the period, the system will be able to will be to fully serve both the data in the buffer and the required traffic of the period.

Figure 12: System (1) Analyzed over 1 Period

Comparison of PNP 2 with Network Calculus

Results

Using the same testbed, traffic production software, and traffic measurement software described in section 3.1.3, we were able to measure the transmitted traffic profile, the received traffic profile, the latency experienced by the data, and the buffer requirements of the transmitter. Based on results from our published work, in which our methods predicted a buffer size of 64,000 bits, we show that Network Calculus predicts a required buffer size of 3155,000 bits. This drastic difference stems from the aforementioned mismatch between downtime and maximum data production.

Network Calculus lacks this capability because it defines its models as functions of time window size rather than as direct functions of time.

Figure 16: System profile used for comparison of PNP 2 with Network Calculus. The analysis using PNP 2 is shown on the right.

Analysis of TDMA Scheduling

Problem
Results

TDMA transmission scheduling affects the timing characteristics of application network communications. For this system, the network between these satellites is a valuable resource shared by each of the application components in the cluster. To ensure the stability of network resources, each satellite has a direct link with all other satellites and is assigned a slot in the TDMA schedule during which the satellite can transmit.

The requirement for accurate performance prediction necessitates the incorporation of the TDMA scheme into the network modeling and analysis. The addition of the TDMA scheduling can affect the buffer and delay calculations based on the slot bandwidth, the number of slots, and the slot length.

Figure 19: Effects of TDMA scheduling in the MAC layer on system network perfor- perfor-mance.

Compositional Analysis

This priority relationship for composition analysis is similar to the task priority used for scheduling analysis in Real-Time Calculus, mentioned in Section 2.1.2.3. Just like RTC, this priority relationship and composition allows us to capture the effects that independent profiles have on each other when they share the same network resources. Just as RTC based its priority relationship and computational scheduling on a fixed-priority scheduler, our priority relationship and resource allocation are based on the network quality-of-service (QoS) management provided by different types of network infrastructure.

One such mechanism for implementing this type of priority-based network resource allocation is the use of DiffServ Code Point (DSCP)[41]. The queue of the lowest priority profiles is considered when the lowest priority profile correlates with the remaining capacity that the node has available to serve it.

Delay Analysis

Analysis of Statically Routed Networks

Problem
Results

The full description of the OOB channel and how the receiver limits the sender can be found in Section 4.2. First, formalizations for semantics and modeling and analysis techniques were defined, based on the (∧,+)-calculus used by the Calculus Network. Finally, we extended our traffic producer/consumer code to enable management of network traffic by the communications middleware.

In a similar way to modeling uncertainty analysis, a temporal uncertainty analysis could be performed to determine the effects of imperfect temporal synchronization between system nodes. Stochastic processes appearing in queuing theory and their analysis using the nested Markov chain method. Annals of Mathematical Statistics.

Figure 20: Experimental setup to validate routing, delay, and compositional analysis of network profiles.

Run-Time Network Performance Monitoring and Management for Dis-

Middleware-Integrated Measurement, Detection, and Enforcement 73

Results

Using this component modeling framework and the corresponding code generation tools we have developed, the application developer only needs to provide the business logic code for the application; the rest of the middleware and component configuration code is automatically provided by our library. Since the sender's middleware code automatically measures and records egress traffic from the application, we have implemented additional code that can optionally push to the application throwing an exception when the application produces more data than specified in its profile. If such a push occurs, the application is notified and the data is not transmitted over the network.

Similarly, because the receiver middleware code automatically measures and records incoming traffic from each of its senders, we introduced an additional communication channel used by the sender and receiver middleware to enable out-of-band communication that is invisible to Login. For the data in this figure, each message was recorded as a set of timestamps, the size of the message, where timestamp is the time the application sent the message.

Figure 22: Schematic representation of a software component.

Distributed Denial of Service (DDoS) Detection

Problem
Results

Given the definition of the fundamental operations of PNP2, analysis of periodic systems was presented. The modeling and analysis extensions described above will pave the way for analyzing the effects on application performance caused by these lower-level protocol mechanisms. InProceedings of the IEEE International Conference on Space Mission Challenges for Information Technology, SMC-IT, 2014, Laurel, MD, USA.

In Proceedings of 4th International ACM SIGBED Workshop on Design, Modeling, and Evaluation of Cyber-Physical Systems, CyPhy '14, pages 44–47, New York, NY, USA, 2014. In Proceedings of the {SCS} Conference on Modeling Communication and simulation of distributed networks and systems, pages 3–11, 2001.

Figure 26: Nodes in an example network and how they communicate (using pub- pub-/sub)

Conclusions and Future Work

Future Work

This type of return path modeling and feedback system required for modeling such protocols would also benefit the analysis of data-dependent application profiles, as these are similarly dependent on external inputs that at least partially control the characteristics of the traffic. they produce. If, instead of precise knowledge about the system and application profiles, the application developers and system integrators have uncertainties associated with their models, then analysis of the uncertainty and its effect on the predictions would expand the scope of systems to which the techniques can be applied . InProceedings of the 4th ACM SIGBED International Workshop on Design, Modeling, and Evaluation of Cyber-Physical Systems, CyPhy '14, pages 44-47, New York, NY, USA, 2014.

We configure the system's static routes using the Linux built-in IPRoute[25] tool, which allows for the configuration of the kernel's routing tables, network address translation (NAT), and network interface characteristics such as the maximum transmission unit (MTU ). In Proceedings of the First International Workshop on Data Distribution for Large-Scale Complex Critical Infrastructure, DD4LCCI ’10, pages 23–28, New York, NY, USA, 2010.

Publications

Highly Selective Conference Papers

Other Conference and Workshop Papers

Submitted Papers - Awaiting Reviews

This filtering ensured that high-priority flow packets would be filtered into the high-priority prio qdisc queue, while lower-priority flow packets would be filtered into the low-priority prio qdisc queue. In Proceedings - 17th IEEE International Conference on Embedded and Real-Time Computing Systems and Applications, RTCSA 2011, Volume 1, Pages 295–. Comparison of Real-Time Calculus with Existing Analytical Approaches for Performance Evaluation of Network Interfaces.

Rosmod: A toolkit for modeling, generating, deploying, and managing real-time distributed component-based software using rosmod. Proceedings - Eighth IEEE International Symposium on Distributed Simulation and Real-Time Applications, DS-RT 2004, pp.