Triple Pattern Fragments (TPFs) is a low-cost interface for querying knowledge graphs on the web. In this work, we present teepee, a visualization tool for analyzing the pro ling results of TPF performance. Moreover, teepee allows for comparing the performance of di erent congurations of TPF servers. Attendees will observe how the execution of requests against TPFs is impacted by i) the type and cardinality of the requested triple patterns, and ii) network delays and server workload. The demo is available at http://km.aifb.kit.edu/services/teepee/.
{ The online demo of teepee1, which includes a basic setup of the TPF Pro ler with a limited con guration for the sake of performance. { The source code of teepee, which is available as part of the TPF Pro ler repository2 and can be adjusted and executed locally.
teepee
TPF Profiler Sampling
Triple Pattern
Generation
Response Time Measurement
View
Results
Compare
Results
teepee includes three components for visualizing the performance results of TPFs
(cf. Figure 1): TPF Pro ler, View Results, and Compare Results.
TPF Pro ler [
The attendees of the demo will learn how teepee can be used to gain insights into TPF server performance using its visualization capabilities. The two use cases can be followed using the online version of teepee1 which includes the examples used in this demonstration. In the online demo, the users are encouraged to generate results themselves by con guring and running the TPF Pro ler. Impact of triple pattern properties on TPF performance. First, we select View Results to retrieve a list of all available results generated by the TPF Pro ler. In the list, the examples are listed rst. User-generated results are listed below (indicated by an ID starting with \U"). Next, we select View Results of Example 1. The view shows metadata on the top, indicating that the results are for the DBLP KG with a sample size of m = 100, page size of 100, HDT (a) Response Time Visualization (b) Sample Visualization backend and a remote server. Below, the visualizations show the response times for analyzing di erent factors. As shown in Figure 2a, the response times are visualized as boxplots for the di erent triple pattern types. Hovering over the boxplot shows detailed information of the results. In the example, we can observe, that the triple pattern type has an impact on the response time as the boxplots vary in shape and position. The pattern type hv; r; vi shows the highest median response time. Moreover, the results are visualized according to the answer cardinalities of the triple patterns. The results of the example show a slight increase in the response time for larger answer cardinalities. Additionally, it can be observed that there are no triple patterns with an answer cardinality within the range of 1 104 and 1 106 answers. Most triple patterns show a cardinality below 50 which may allow for optimizing the server performance by reducing its page size. This assumption can be analyzed further using the A/B comparison visualizations presented in the next use case. Furthermore, information on the sampled triples by the TPF Pro ler is provided as well. The number of samples is listed as well as the number of the unique triple patterns generated from the sample. Moreover, the number of distinct subjects, predicates, and objects in the triple patterns are provided allowing to gain an insight into the RDF term distribution. Shown in Figure 2b, all terms in the sample are listed and connected if they commonly occur in the same triple. Furthermore, links for all positions in the triples are added. Hovering over \Subjects" for instance will indicate all terms that occur as subjects in the sample. In the example, the term foaf:maker is selected and it can be observed that it occurs in a large number of triples within the sample.
(a) Boxplot A/B comparison (b) Radar plot A/B comparison Impact of di erent servers via A/B comparison. Second, we select Compare Results to visualize the results in an A/B comparison to gain insights into how di erent TPF server con gurations a ect response times. Therefore, we select Example 2 and Example 6, which are the results for the DBpedia KG di ering in the TPF server environment. The server in Example 2 is a publicly available TPF server hosted remotely at http://data.linkeddatafragments.org. The server in Example 6 is hosted locally. In Figure 3a, the response times for a selection of triple pattern types are shown with results for the remote server on the left and for the local server on the right. Two major insights can be gained from this visualization. First, we observe that the response time for the remote server is 10 times higher than for the local server. This indicates that network delays and server workload are in uencing the response time for the remote server. Second, the response time for the pattern type hv; v; vi is lower with respect to the other pattern types on the remote server. This result suggests that the publicly available server caches the results for this triple pattern type. This is probably due to the fact that hv; v; vi is requested frequently, e.g. each time the web page of the TPF server is loaded. The radar plot in Figure 3b supports the comparison by visualizing various statistics of the response time. 4
This demo presents teepee, a tool that combines the TPF Pro ler with visualization capabilities to enable analyses of TPF performance. teepee allows for gaining a better understanding of how di erent factors impact on the performance of TPF servers. In two use cases, the attendees are able to analyze and compare empirical results over well-known KGs using teepee.