GCLViz: Garbage Collection vs. Latency Visualization

Chihua Ma

, Stanislav Liberman

and Haifeng Zheng

Department of Computer Science, University of Illinois at Chicago, Chicago, U.S.A.

Technology & Enterprise Computing Division, CME Group, Chicago, U.S.A.

Keywords: Information Visualization, Time Series Data Visualization, Detail View, Gc Visualization, Latency,

Performance, Trading System.

Abstract: This paper proposes a method that creates a multi-view interactive visualization that allows users to explore

connections between garbage collection (GC) generated by Java Virtual Machine (JVM) and latency in

applications used in financial transactions. With this tool users can explore large collections of GC and

latency events, easily identify important events, and subsequently focus on the relationships and details of

such events without losing the “big picture” perspective on the events as a whole. We discuss the impact of

this tool on controlling the effects of GC on latency and variability in financial trades with an exchange.

1 INTRODUCTION

Securities exchanges have increasingly adopted a

limit order market design, in which traders submit

orders directly into exchange’s electronic systems,

bypassing both designated and unofficial market

makers. This transition from traditional to fully

electronic limit order market (Kirilenko and Kyle,

2011) has made the rigid distinction between market

makers and conventional traders obsolete. This

transformation has occurred due to advances in

technology, as well as regulatory requirements.

An increasingly important dimension of

electronic trading is latency (Brogaard, 2010): the

time between the release of a market message by the

participant and its reception and execution by the

exchange computers. Another important factor is

variability. Variability in this context refers to

dissimilarity in response times observed by market

participants. Many factors may introduce variability

into the message flow: network latency, packet

retransmissions, operating system (OS) network

stack and OS task scheduling, or the application

itself. In order to reduce both latency and variability

“co-location” is typically used, whereby market

participants will rent space in a computer server

centre next to an exchange. This approach leverages

physical proximity to reduce the time a market

message takes to arrive at the exchange. In addition,

at such centres engineers work to optimize high-

performance trading by eliminating the factors

causing the latency and variability. One important

factor that may adversely affect performance is

garbage collection (GC) generated by Java virtual

machine (JVM).

Use of managed runtime-based languages in high

performance computing environment is a fairly new

development. For instance, recent advances in Java’s

technology and techniques have made it a

predominant platform in low latency applications

(Lawrey et al., 2013). Java and other managed

runtimes have been widely used in production for a

number of significant application areas, including

financial trading, telecommunications, and military

command-and-control (Auerbach et al., 2008).

Companies find Java attractive due to vast array of

libraries, frameworks, tools, IDEs, and server

providers; and it runs on a variety of platforms and

CPU architectures.

JVM is the code execution component of the

Java platform. The specification dictates that any

JVM implementation must include the automatic

memory management service – known as GC. Java

Memory Management (Reitbauer et al., 2011), with

its built-in GC, allows developers to create new

objects without worrying explicitly about the

memory allocation and deallocation, because the

garbage collector automatically reclaims memory for

reuse. This enables faster development with less

boilerplate code, while eliminating memory leaks

and other memory-related problems. However, the

behaviour and efficiency of a garbage collector can

292

Ma C., Liberman S. and Zheng H..

GCLViz: Garbage Collection vs. Latency Visualization.

DOI: 10.5220/0004740902920299

In Proceedings of the 5th International Conference on Information Visualization Theory and Applications (IVAPP-2014), pages 292-299

ISBN: 978-989-758-005-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

heavily influence the performance and

responsiveness of any application that relies on it.

Given this link between GC and latency, it would

be beneficial to quickly explore the influence of a

GC-indicated allocation problem on latency in a real

trading system, since time is money in the exchange

business. However, examining the text in the

massive data files produced by the GC and latency

logging outputs can be a daunting task for engineers.

What is needed is a powerful, effective graphical

presentation of those data, since graphs can convey

such time-series information more quickly and more

informatively than just text to the human operator.

In this paper we present the design of Garbage

Collection vs. Latency Visualization [GCLViz], a

user interactive visualization tool that turns both GC

and latency logging outputs into x/y time-series line

plots, bar charts, and other unique types of graph,

including a time-overlap-view and a correlation-

circle-view. GCLViz allows the user to load the raw

GC and latency logs directly into its application

which does all the processing behind the scenes.

GCLViz enables users to explore large collections of

GC and latency events, easily identify interesting

events, and subsequently focusing on the relations

and details of such events without losing the “big

picture” (Sekhavat and Hoeber, 2013) perspective on

the collection as a whole.

We first summarize related work in visualization

of GC and system performance, and introduce the

features of our data. We then describe the design of

the GCLViz visualization in detail, followed by

conclusions and future work.

2 RELATED WORK

In recent years, there have been a large number of

real-time dynamic visualizations for software

systems (Reiss 2003). Perhaps the most prominent

effort is IBM’s Jinsight (De Pauw et al., 2001).

Jinsight typically operates by collecting detailed

trace data as the target program executes and when

execution is complete, it uses a variety of views

based on the trace, to allow the programmer to

understand execution at a very detailed level.

JConsole (Java SE Monitoring and Management

Guide, Using JConsole, n.d.) is a graphical

monitoring tool to monitor JVM and java

applications. JConsole provides information on

performance and resource consumption of

applications running on the Java platform using Java

Management Extensions (JMX) technology.

LagAlyzer (Adamoli et al., 2010) is a tool to analyze

and visualize the information of traces produced by

other latency measurement tools. LagAlyzer is an

offline tool: it requires the completed traces to exit

before it can start to analyze and visualize them. Lila

Viewer (Adamoli et al., 2010) is also a visualization

tool that draws trace timelines showing the start and

end of each interactive request. To visualize the

distribution of latencies over time, Analytics (Gregg,

2010) used a heat map created with time on the x-

axis and latency on the y-axis. The heat map is a

colour-shaded matrix of pixels, where each pixel

represents a particular time and latency range.

In addition to visualizing software performance,

there are numerous packages involved in the

visualization or interpretation of garbage collection

data. GCViewer (Schreiber, 2002) is a small tool

that visualizes verbose GC output generated by

Sun/Oracle, IBM, and BEA Java Virtual Machines.

It also calculates garbage collection related

performance metrics (throughput, accumulated

pauses, longest pause, etc.). HPjmeter (Tool Report:

HPjmeter, 2002) is designed to display the collected

metrics to allow the user to easily identify

performance bottlenecks and quickly tune the Java

applications. The IBM Monitoring and Diagnostic

Tools for Java – Garbage Collection and Memory

Visualizer (GCMV) (n.d.) is a tool which allows the

user to visualize and analyze the memory usage and

garbage collection activity of the Java application.

Due to the nature of those tools, this kind of analysis

can only be performed by a small group of expert

users that have high technical skills. To allow a

wider range of testers to carry out expert analysis,

GcLite (Angelopoulos et al., 2012) tool has been

created for analyzing garbage collection logs.

While the tools mentioned above work well as

far as they go, we believe there is still a need for

new designs and techniques. For example, we have

other requirements to consider and therefore

conclude that none of the existing tools met our

needs. To the point, visualization of the impacts of

GCs on latency is needed, and techniques for

incorporating analytical tools within the simplified

domain of end-user visualization would prove

useful. GCLViz can help engineers establish or

refute a correlation between time-of-day-based

observed latency and JVM-wide GC behaviour. Our

visualization helps one to quickly understand the

impact a GC-related change on latency.

GCLViz:GarbageCollectionvs.LatencyVisualization

293

3 DATA

3.1 Garbage Collection Output

In our study, the GC log is the standard output form

(Figure 1) for Oracle JVM. Standard Oracle JVM

1.7.0_25 64-bit was used on servers with Intel CPUs

running Linux operating system. Standard options

for reporting details of each garbage collections

were used. The default garbage collector on this

platform is Parallel Collector (Garbage Collector

Ergonomics, n.d.), which is what was used to run the

simulations. Detailed description of the operation of

this garbage collector is beyond the scope of this

paper, but is readily available from Oracle as well as

descriptions of other available collectors.

Figure 1: Garbage Collection logging output.

In this paper, we only consider minor collections

for young generations. The critical attributes of GC

logging output are the collection time when GC

happens – known as GC event time, and the duration

GC lasts – known as GC event duration. GC event

time includes the year, month, day, hour, minute,

second, and millisecond. GC event duration is the

time period in millisecond.

3.2 Latency Output

In trading-systems, latency is defined as the time the

exchange takes to react to the market. In general,

latency measures the delay between an action and a

response. Like GC log, the most useful attributes of

latency logging output are latency event ID, latency

event time, and latency event duration. Each latency

event has a specific event ID. Latency event time

and duration have the same format as GC.

3.3 Simulation Data

To protect sensitive business operations, we used

simulated data. Consequently, all the figures shown

in this paper are generated using this simulated data.

For the purposes of demonstrating the impact of GC

on application performance in terms of latency, we

have chosen to use Jetty (n.d.) web application

server as the server component. It was chosen

mainly due to its good performance characteristics

as well as ease of configuration. We create a custom

client using Java, which would retrieve JSP page

rendering the HTTP request from the server.

While dynamics of a web application server is

quite different from a matching engine of an

exchange, the fundamental approach of operation is

the same as that of exchange. The request is received

over the network connection, acted upon by the

server, and then a response is sent to the user. For

vast majority of Java applications the processing

stage will result in memory allocations, which will

lead to garbage collection. Our goal is to examine

the impact of those GCs on latency of responses to

requests that were made at that time.

Latency was measured by the client application

with nanosecond precision and each observation was

recorded for analysis.

4 VISUALIZATION DESIGN

By presenting a large number of GC and latency

events in a single view, it is not easy for users to

recognize the relations between GCs and latencies.

We design GCLViz to follow Shneiderman’s visual

Information Seeking Mantra: “overview first, zoom

and filter, then details-on-demand” (Shneiderman,

1999). In other words, in the exploratory data

analysis (EDA) of a data set, an analyst first obtains

an overview. This may reveal potentially interesting

patterns or certain subsets of the data that deserve

further investigation. The analyst then focuses on

one or more of these, inspecting the details of the

data. The goal of GCLViz is to provide an effective

visual representation that scales well with a large

number of events and allows the user to explore the

relation between GC and latency at a micro-level.

4.1 Overview

The three main synchronized parts of GCLViz,

shown in Figure 2, are the global-view, detail views

and a table-based view. The global-view provides an

overview of all of the events. The detail views which

includes a time-overlap-view, a correlation-circle-

view and scatter plots, shows the subset of events

selected from the global-view, clearly illustrating the

relationships between GC and latency. Interesting

events can be accessed through the table-based view

and highlighted within the global and detail views.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

294

Figure 2: A screenshot of the GCLViz system. The global-view of the entire set of events on the left, the detail views,

including time overlap, correlation circle and scatter plot, of the events related to a selection of items on the right, and a

table-based view of interesting events on the bottom-right.

4.2 Global-view

The global-view in GCLViz is a 2D graph

representation for time series data with line graphs

and bar charts that provides a “big picture”

(Sekhavat and Hoeber, 2013) overview of both GC

and latency events and allows users to identify

interesting aspects within data. Multiple line graphs

or bar charts can be used or overlaid to show more

than two dimensions (x, y1, y2 …) (Kromesch and

Juhász, n.d.). In GCLViz, we use different colour

coding to distinguish each dimension. Each

dimension may be drawn using a different scale.

In global-view, there are seven dimensions: x

value always represents time variable, while six y

values are used to represent internal latency

duration, internal latency count (frequency of

occurrence), external latency duration, external

latency count, GC duration, and GC count. We use

three sets of channels: Yellow-green series are used

for encoding internal latency events, pink series are

used for external latency events, and blue series are

used for GC events. Dots represent GC and latency

events. Line graphs show the “event duration” while

bar charts present the “event count” over entire

timeline (24 hours). In addition, due to the large size

of the external latency data (about 200 events in a

second), the original data is down-sampled by a

factor of 10 when time period on the x-axis is larger

than one minute.

Interaction. Interaction is an important element in

any visualization system. Providing an interaction

for users to explore the data helps them to perceive

the relations within data. In GCLViz, basic

interaction is done with simple mouse and keyboard

operations. Users can select an event they wish to

investigate further by clicking on the data displayed

in the global-view.

When users select a specific data point in the

global-view, the corresponding event is highlighted

by a semi-transparent grey bar with a circle

displaying the event duration. Both the time-overlap-

view and correlation-circle-view are updated to

reflect the corresponding time when the event

happens, and the details that are related to that event

are shown in the information panel in the time-

overlap-view, as illustrated in Figure 2. Manual

zooming is performed by dragging the time sliders

of hour, minute, and second at the bottom of the

graphs. Figure 2 displays the graphs during the time

from 17:0:0 to 17:59:59. User can also enable or

hide any one or multiple graphs by checking or

unchecking the checkboxes on the top of the global-

view. For example, Figure 2 only shows the graphs

GCLViz:GarbageCollectionvs.LatencyVisualization

295

of external latency duration, GC duration and GC

count.

Once users select events from the global-view that

they deem important, they can further explore the

relationships among these events, such as the latency

related to a specific GC and inspect details of

selected events, with the time-overlap-view. In the

time-overlap-view, we use both X-axis and Y-axis to

represent time and duration as well. All the events

start on the line where y = x, and end at the point (x

+ d, x), where d is the event duration. The durations

of events are drawn along x-axis by using the lines

with same color-code as the one in the global-view.

The end points of events are drawn by using dots.

Where the global-view displays the performance

over a twenty-four hour period; the time-overlap-

view provides a snapshot over a second as default.

This view can be magnified 1000 times to one

millisecond.

Figure 3 shows the time-overlap-view expanded

to 30 milliseconds after the user clicks an event in

the global-view. The chosen event is highlighted

with its detail information displayed in the

information panel. We can obviously see that most

of the events of external latency coloured with pink

that concentrate on the diagonal line have very short

durations (within nanoseconds). The only long

external latency starts at 9:3:42:931 and lasts 5.15

milliseconds. It happens right before a GC event

coloured with blue. This figure strongly

demonstrates the impact of GC on latency.

Interaction. Clicking a data point causes the

corresponding event to be highlighted with a semi-

transparent grey bar consisting of a rectangle topped

by a triangle. It highlights all the events that overlap

with the event selected. Meanwhile, the information

panel at the top-left of the time-overlap-view

displays the detail information of the chosen event,

including event ID, event starting time and event

duration. The user can look into the event in other

relative log files according to the event ID.

Pressing “+/-” keys on the keyboard will zoom

in/out. The starting point of the chosen event is

always centred in the time-overlap-view. It is

possible that the user cannot see the complete event

when the duration is longer than half of the time

period of current view. The “arrow” buttons on both

sides of the x-axis are used for shifting the whole

overlap view one millisecond forward or backward

to place the chosen event in a proper location. As the

time sliders in the global-view, the time sliders here

can allow the users to make the time-overlap-view

jump to a particular time point directly by dragging

them to that time.

Figure 3: Time-overlap-view when zooming in.

4.4 Correlation-Circle-View

The correlation-circle-view borrows the idea of

Circle View (Keim, 2004) technique, which is a

combination of hierarchical visualization techniques

(Shneiderman, 1992), such as tree maps and circular

layout techniques (Ankerst, 1996), such as Pie

Charts and Circle Segments. The main goal is to

compare continuous data over time in a limited

display space, in order to identify patterns,

exceptions and similarities in the data. The basic

idea of the correlation-circle-view display is to

visualize the change of correlation between GC

count and latency count over time.

We have two ways of counting the events: counts

per minute and counts per second. In this paper, we

only discuss the case when the events are counted

per second. For example, we have sixty samples for

one minute of data. We applied Pearson’s

correlation coefficient to the pair of the series of

numbers. The window size (length of time) for

calculating the linear correlation coefficient is set to

40 seconds as default. The correlations are

calculated from the beginning of both events of

latency and GC then shifted one second for each step

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

296

over the entire timeline to generate a correlation

coefficient value. To display a series of correlation

coefficient values over time in a limited display

space and provide users an effective view to observe

how the event count changes in a minute or hour, we

choose the circle view.

Figure 4: Correlation-circle-view.

Figure 4 shows an example structure of a single

circle view dividing a circle in 60 segments. Each

segment represents a correlation coefficient value.

Both the lengths of fans and opacity factors are

determined by the correlation coefficient values. Red

fans represent positive correlation values, while

green ones represent negative correlations. For

example, Figure 4 shows the change of correlation

between GC count and latency count during the 53

minute at 5 pm with window size of 20 seconds. The

highlighting green fan presents that the correlation

value for the series of event count starting from

17:52:50 to 17:53:10 is -0.229.

From a perceptual point of view, it is easier to

compare segments, which are located very close to

each other. The eye of the data analyst can directly

compare neighbouring time slots or even

unconnected time slots.

Interaction. The arcs of segments can be

investigated as well. For example, clicking the 21

segment in Figure 5 causes it to be highlighted by a

semi-transparent grey fan with the time sliders

jumping to 17:52:21. Dragging the minute time

slider also causes the circle view to be updated to the

corresponding minute. The visualization works for

different window sizes by pressing “+/-” key. In

Figure 4 and 5, the window size is decreased to 20

seconds from the default value (40 seconds) by the

user. The correlation-circle-view must be updated

each time new parameter values are assigned. In

addition, other views are updated as well. A semi-

transparent red bar is drawn at 17:52:21 with the

width of window size (20 seconds) in the global-

view (Figure 5). The time period for displaying in

the time-overlap-view is updated to from 17:52:21 to

17:52:22.

Figure 5: A correlation coefficient value in the

Correlation-circle-view and its relative position in the

Global-view.

4.5 Scatter Plot

The scatter plot is another kind of correlation. The

more the data sets agree, the more the data tend to

concentrate in the vicinity of the identity line. Each

semi-transparent yellow dot represents a particular

period of one second. Figure 6 presents 3600 dots

for the period from 5 pm to 6 pm. During the period

of one second, we count the number of events and

calculate the average and maximum duration of

events.

A Grid of two-dimensional scatterplots is the

standard way of extending the scatter plot to higher

dimensions (Kromesch and Juhász, n.d.). Since each

data set has three dimensional data, a three by three

array of scatter plots is used to provide a

visualization of each dimension versus every other

dimension. This is useful for looking at all possible

two-way interactions or correlations between

dimensions. The nine scatter plots are GC count vs.

latency count, GC count vs. latency Avg duration,

GC count vs. latency Max duration, GC Avg

duration vs. latency count, GC Avg duration vs.

latency Avg duration, GC Avg duration vs. latency

Max duration, and GC Max duration vs. latency

GCLViz:GarbageCollectionvs.LatencyVisualization

297

count, GC Max duration vs. latency Avg duration,

GC Max duration vs. latency Max duration. There

are three by three rectangles in a table for the users

to select one of nine scatter plots by pressing the left

mouse button inside of the corresponding rectangle.

Figure 6 shows the relation between GC average

duration and latency average duration in every

second.

Figure 6: One of the nice Scatter Plots.

Figure 7: Table-based View.

4.6 Table-based View

The table-based view (Sekhavat and Hoeber, 2013)

is useful for users to find interesting events. By

comparing GC and latency events, the table provides

top five interesting events for each of the five

categories: Latency Duration, GC Duration, Number

of Overlaps between GCs and Latencies, Number of

Latencies per Second, and Number of Latencies per

GC. Figure 7 indicates top five GC events with

largest number of latencies between them and the

following GCs. The table-based view is clickable.

All the views will be updated when users click the

event time in the table.

5 CONCLUSIONS AND FUTURE

WORK

In this paper we propose a new technique for the

visualization of GC vs. Latency. The goal of

GCLViz is to build a visualization system for the

exploration and analysis of the impact of GC on

latency and variability. Using simulated data rather

than actual data from the exchange company to

protect sensitive business operations, we

demonstrate that GCLViz does provide important

visualizations to show GC impacts on latencies. The

most important value of GCLViz is in its ability to

provide engineers information that can be used to

control and minimize the effects of GCs impacts on

system performance.

GCLViz was designed to follow Shneiderman’s

visual information seeking mantra of “overview

first, zoom and filter, then details on demand”.

GCLViz presents the GC and latency in multiple

synchronized views. The scale of displayed

information and layout were chosen to support

observed behaviour and allow users to expand

visualized data at a micro-level in detail views

including the time-overlap-view, the correlation-

circle-view and the scatter plots, that illustrate

relationships and further details on a subset of the

events, while still providing the relative position of

the subset of events in an overview. Interactive

highlighting makes exploring the events selected in

different views an effortless process. Finally, the

table-based view consisting of “interesting events” is

provided to help users find critical events quickly.

While there are a great number of sophisticated

tools available that focus on visualizing GC or

system performance, GCLViz offers two main

advantages over other tools in the field. It builds the

connections between GC and latency allowing users

to explore relationships between them; it develops a

2D time-overlap-view for visualizing data of which

both dimensions are time variables.

In the future, it would be very meaningful to

explore the behaviours of GCs by using different

type of collectors and compare their impacts on

latency. Besides the default garbage collector

(Parallel Collector) we use in our current method,

there are three additional collectors (Java SE 6

HotSpot Virtual Machine Garbage Collection

Tuning, n.d.): Serial Collector, Concurrent

Collector, and Garbage-First Collector. Each of

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

298

them is a generational collector which has been

implemented to emphasize the throughput of the

application or low garbage collection pause times. In

addition, there are different kinds of measurement of

latency as well. Visualizing multiple data sets in the

same view can be a big challenge just like

visualizing high dimensional data. We need to

develop more efficient visual layouts for visualizing

the large-scale time-based data. Finally, we plan to

integrate the data online searching into this

application tool rather than analysing the local data

offline, so that the visualization can be implemented

in real-time.

ACKNOWLEDGEMENTS

We would like to thank Professor Robert V. Kenyon

for his help in editing the paper and his advice on

how to organize the paper.

REFERENCES

Adamoli A., Jovic M., and Hauswirth M., 2010.

LagAlyzer: A Latency profile analysis and

Visualization tool. In ISPASS ’10, Proceedings of the

2010 IEEE International Symposium on Performance

Analysis of System and Software. IEEE.

Angelopoulos V., Parsons T., Murphy J., and O’Sullivan

P., 2012. GcLite: An Expert Tool for Analyzing

Garbage Collection Behavior. Proceedings of the 2012

Computer Software and Applications Conference

Workshops. 2012 IEEE 36

Annual, pp. 493-502.

Ankerst M., Keim D. A., and Kriegel H. –P., 1996. Circle

segments: A technique for visually exploring large

multidimensional data sets. In Visualization ’96, Hot

Topic Session, San Francisco, CA.

Auerbach J., Bacon D. F., Cheng P., Grove D., Biron B.,

Gracie C., McCloskey B., Micic A., and Sciampacone

R., 2008. Tax-and-Spend: Democratic Scheduling for

Real-time Garbage Collection. EMSOFT ’08,

Proceedings of the 8

ACM International Conference

on Embedded Software, Atlanta, GA, USA.

Brogaard J. A., 2010. High Frequency Trading and Its

Impact on Market Quality. Ph.D. Thesis. Northwestern

University, USA.

De Pauw W., Mitchell N., Robillard M., Sevitsky G., and

Srinivasan H., 2001. Drive-by analysis of running

programs. Proceedings of ICSE Workshop of Software

Visualization, International Conference on Software

Engineering, Toronto, Ontario, May 2001.

Garbage Collector Ergonomics, 2013. Retrieved from:

http://docs.oracle.com/javase/7/docs/technotes/guides/

vm/gc-ergonomics.html

Gregg B., 2010. Visualizing System Latency.

Communications of the ACM. vol. 53, no. 7, pp. 48-54.

IBM Monitoring and Diagnostic Tools for Java – Garbage

Collection and Memory Visualizer, 2013. Retrieved

from:

http://www.ibm.com/developerworks/java/jdk/tools/gc

mv/

Java SE 6 HotSpot Virtual Machine Garbage Collection

Tuning, 2013. Retrieved from:

http://www.oracle.com/technetwork/java/javase/gc-

tuning-6-140523.html#available_collectors

Java SE Monitoring and Management Guide, Using

JConsole, 2013. Retrieved from:

http://docs.oracle.com/javase/6/docs/technotes/guides/

management/jconsole.html

Jetty, 2013. Retrieved from: http://www.eclipse.org/jetty/

Keim D. A., Schneidewind J., and Sips M., 2004.

CircleView – A New Approach for Visualizing Time-

related Multidimensional Data Sets. In ACM Advanced

Visual Interfaces (AVI). Association for Computing

Machinery (ACM). ACM Press.

Kirilenko A. and Kyle A. S., 2011. The Flash Crash: The

Impact of High Frequency Trading on an Electronic

Market. Manuscript, U of Maryland, USA.

Kromesch S and Juhász S., 2013. High Dimensional Data

Visualization. Retrieved from:

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.

1.1.108.3671

Lawrey P., Thompson M., Montgomery T. L., and Piper

A., 2013. Virtual Panel: Using Java in Low Latency

Environments. Retrieved from:

http://www.infoq.com/articles/low-latency-vp

Reiss S. P., 2003. Visualizing Java in Action. SoftVis ’03,

Proceedings of the 2003 ACM Symposium on Software

Visualization. pp. 57-65.

Reitbauer A., Enzenhofer K., Grabner A., Kopp M.,

Pierzchala S., and Wilson S, 2011. Java Enterprise

Performance, Compuware Corporation. Retrieved

from:

http://javabook.compuware.com/content/start.aspx

Tool Report: HPjmeter, 2002. Retrieved from:

http://www.javaperformancetuning.com/tools/hpjmete

r/index.shtml

Schreiber H., 2002. GCViewer. Retrieved from:

http://www.javaperformancetuning.com/tools/gcviewe

r/index.shtml

Sekhavat Y. A. and Hoeber O., 2013. Visualizing

Association Rules Using Linked Matrix, Graph, and

Detail Views. International Journal of Intelligence

Science, 3, 34-49.

Shneiderman B., 1999. The Eyes Have It: A Task by Data

Type Taxonomy for Information Visualizations.

Proceedings of the IEEE Symposium on Visual

Languages. Boulder, pp. 336-343.

Shneiderman, B., 1992. Tree visualization with tree-maps:

2-d space-filling approach. ACM Transactions on

Graphics (TOG), pp. 92–99.

GCLViz:GarbageCollectionvs.LatencyVisualization

299