Citation Context

...ponents are responsible for causing faults. Our work focuses on recovering the task trees associated with multi-layer protocols, rather than the analysis of those recovered paths. Aguilera et al., in =-=[1]-=-, find anomalous behavior in distributed systems by treating each component as a black box, and inferring the operation paths by only looking at message traces. They present heuristics to recover the ...

Citation Context

... for tracing large-scale operations through the network. It is more useful for tuning small blocks of code than distributed algorithms and their emergent behavior. More recent tools such as Project 5 =-=[1]-=-, Magpie [2], and Pinpoint [5] infer causal paths based on traces of netIn Proc. 3rd Symp. on Networked Systems Design and Implementation (NSDI), San Jose, CA, May, 2006 2 Approach Scenario gdb and g...

Citation Context

...e entry. 4) On the measure of root-cause ranking, PeerPressure’s yields better ranking for most of the cases. Another interesting work that also takes the blackbox approach is that of Aguilera et al. =-=[1]-=-. They address the problem of black-box performance debugging for distributed systems. They developed and compared two algorithms for inferring the dominant causal paths. One uses the timing informati...

Citation Context

...lation will provide ground truth against which we can compare the results of the clustering and retrieval operations. 4.1 Claim 1: Evaluating Proposals for Signatures We evaluate and compare three possible approaches for creating signatures. Let ~S denote the vector representing the signature. In all cases the elements si in this vector correspond to a specific system, application, or workload metric. 1. Raw values: in this case we represent a signature ~S using the raw values of the metrics. In other words ~S = ~M . Following common practice in data analysis [6], we normalize these values to [0, 1], using the range of values ever seen, to prevent scaling issues from influencing similarity metrics and clustering.5 This signature is the most naive and requires no extra processing of the traces. 2. Metric attribution: If the attribution as described in Section 2.2 flags the metric mi to be attributable, then si = 1, otherwise si = −1. If the metric mi is not even considered by the models for the attribution process, then the metric is considered to be irrelevant to the SLO state, and si = 0. This requires significant computation [5], but we have shown that the computations could be done on...

Citation Context

...ssed to infer the relevant properties. Examples of this approach are Magpie for performance modeling in clusters [5], Aguilera et al.’s trace analysis for performance debugging in distributed systems =-=[1]-=-, and SPIE for identifying the source of a packet in the Internet [46]. Summary logs such as those collected using Cisco’s NetFlow [9] have proven useful for network operators to diagnose problems wit...

Citation Context

...s returned by our system may be more relevant because they pertain exclusively to the system under consideration. Several recent projects have investigated path-based debugging of distributed systems =-=[10, 1]-=-. These systems log the interactions between components or nodes of a distributed system. By applying statistical techniques to these traces, it is possible to extract some information of interest, e....