On Effectiveness of Compressed File Transfers to/from the Cloud:

An Experimental Evaluation

Armen Dzhagaryan and Aleksandar Milenkovi

Electrical and Computer Engineering Department, The University of Alabama in Huntsville, Huntsville, AL, U.S.A.

Keywords:

Cloud Computing, Lossless Compression, Energy Efﬁciency, Performance Evaluation, Cost Efﬁciency.

Abstract:

An increasing reliance on cloud and distributed processing of scientiﬁc and big data in commercial, academic,

and government institutions necessitate new approaches to optimize ﬁle transfers. Lossless data compression

and decompression is essential in improving the overall effectiveness of ﬁle transfers between edge devices

and the cloud by increasing communication throughput, reducing connection latency, making effective use of

cloud storage, and reducing costs. This paper experimentally evaluates effectiveness of common and emerging

general-purpose compression utilities for ﬁle transfers between edge devices and the cloud. The utilities are

evaluated in terms of throughput and costs during representative ﬁle transfers between a workstation and the

cloud, while varying LAN network conditions. The results show that the optimal compressed transfer modes

improve both upload and download throughputs. For uploads, the peak improvements range from 5.16 to 25.6

times relative to uncompressed ﬁle uploads, and from 1.33 to 17.4 times relative to the default compressed

uploads. For downloads, the peak improvements range from 3.82 to 19.57 times relative to uncompressed

downloads, and from 1.8 to 13.8 times relative to the default compressed downloads. In addition, the best

performing compressed transfer modes reduce the costs related to cloud computing.

1 INTRODUCTION

Cloud and distributed computing are used by com-

mercial, academic, and government institutions to

process and analyze big data generated from multi-

ple sources (Chen et al., 2014). For example, big data

applications may analyze web or users online activi-

ties generated from public web and social media to

make predictions in areas such as, product evaluation

and market characterization. An increased collabora-

tion and data exchange in scientiﬁc communities re-

sult in an increased communication between edge de-

vices and the cloud. Finally, emerging machine lear-

ning applications require transfer of a large amount of

training data to and from the cloud.

The cloud computing poses new challenges asso-

ciated with costs, security, and storage (Abadi, 2009;

Abu-Libdeh et al., 2010; Sboner et al., 2011). Whe-

reas the use of cloud services opens up new oppor-

tunities in computing and reduces the costs of ow-

ning, operating, and maintaining hardware, the costs

associated with the use of cloud services can even-

tually become prohibitively high for small research

and industry organizations. Providers of cloud plat-

forms charge utilization fees for using computing re-

sources and data transfer fees for data transfers either

to or from the cloud (Microsoft, 2017; Amazon, 2017;

Google, 2017a). The speciﬁc cloud instance conﬁgu-

ration (disk space, tenancy type, network priority, and

computational power) and location of the cloud in-

stance determine the ﬁnal utilization and transfer fees

associated with each instance type. Consequently, op-

timizing data transfers between edge devices and the

cloud is very important for a range of cloud compu-

ting applications.

The importance of lossless compression in opti-

mizing throughputs and costs in cloud computing has

been recognized by both industry and academia. Los-

sless data compression is currently being used to re-

duce the required bandwidth during ﬁle downloads

(e.g., software repositories, distributed storage) and to

speed up web page loads in browsers. Several studies

showed that lossless compressed transfers between

edge devices and the cloud can increase throughput

and reduce overall cloud costs during transfers (Sbo-

ner et al., 2011; Nicolae, 2010; Bonﬁeld and Maho-

ney, 2013; Bicer et al., 2013). Another study is focu-

sed on tradeoffs of compression in reducing time and

energy of IO in Hadoop (Chen et al., 2010). Studies

focusing on transfers over WLAN and cellular inter-

Dzhagaryan, A. and Milenkovi

c, A.

On Effectiveness of Compressed File Transfers to/from the Cloud: An Experimental Evaluation.

DOI: 10.5220/0006905800350046

In Proceedings of the 8th International Joint Conference on Pervasive and Embedded Computing and Communication Systems (PECCS 2018), pages 35-46

ISBN: 978-989-758-322-3

faces on mobile platforms showed that not a single

combination of a compression utility and a compres-

sion level performs the best for all ﬁle transfers and

network conditions (Barr and Asanovi

c, 2003; Barr

and Asanovi

c, 2006; Milenkovi

c et al., 2013; Dz-

hagaryan et al., 2013; Dzhagaryan and Milenkovi

2015; Dzhagaryan et al., 2015). Using measurement-

based observations, a couple of studies introduce run-

time techniques for deciding whether to use compres-

sed transfer or not (Krintz and Sucu, 2006; Nico-

lae, 2010; Harnik et al., 2013). Our previous works

on embedded and mobile devices introduced analy-

tical models for estimating effective throughput and

energy efﬁciency for uncompressed and compressed

ﬁle transfers (Dzhagaryan and Milenkovi

c, 2016), as

well as a run-time technique using those models (Dz-

hagaryan and Milenkovi

c, 2017). The importance

of lossless data compression is further underscored

with the recent development of new compression al-

gorithms, such as Apples lzfse (Apple, 2017), Face-

books zstd (Facebook, 2017), and Googles brotli (?).

All three are designed to replace the existing compres-

sion algorithms, such as zlib, by employing newer en-

coding methods (e.g., FSE (Cyan, 2013) in zstd and

lzfse) and focusing on optimizations to improve per-

formance and energy efﬁciency. Three compression

utilities were designed speciﬁcally for mobile devi-

ces and application repositories (lzfse), data centers

(zstd), and web compression (brotli).

In this paper, we perform a comparative,

measurement-based study of general-purpose com-

pression utilities with the focus on performance and

cost effectiveness in ﬁle transfers between a worksta-

tion and the cloud. We perform compressed ﬁle trans-

fers to/from the cloud instances residing in North Vir-

ginia and Tokyo for a range of input ﬁles and network

conditions. Measured throughputs and data transfer

cloud costs of each selected utility are compared to

the throughputs and costs of uncompressed ﬁle trans-

fers and the default compressed ﬁle transfers that use

gzip with -6 compression level.

We ﬁnd that the compressed ﬁle transfers signi-

ﬁcantly improve effective throughputs and reduce the

cloud costs. The level of reduction depends on (i) cha-

racteristics of an input ﬁle; (ii) network connection;

(iii) cloud pricing for data transfers and selected in-

stance type and location, and (iv) performance cha-

racteristics of the workstation that initiates the trans-

fer and the cloud instance performing compression.

Depending on the ﬁle type, selecting optimal com-

pressed uploads improves effective throughputs up to

5.16 25.6 times relative to raw uploads, and 1.33 17.4

times relative to the default compressed uploads with

gzip -6. Optimal compressed downloads improve do-

Figure 1: File uploads and downloads in cloud computing.

wnload throughputs up to 3.82 29.57 times relative

to raw downloads, and 1.8 13.8 times relative to the

default downloads with gzip -6.

We ﬁnd that optimal compressed ﬁle upload mo-

des reduce the costs associated with data transfers up

to 83 - 99.6% relative to uncompressed uploads and

up to 25.4 - 90.3% relative to the default compressed

uploads. Optimal compressed ﬁle download modes

lower the download costs relative to uncompressed

downloads up to 86.1 - 93%, and up to 42.9 - 83.7%

relative to the default compressed downloads.

Optimal data uploads that improve effective

throughputs and costs are achieved using pzstd/zstd

on connections with higher network throughputs, and

using lbzip2/pbzip2 on connections with lower net-

work throughputs. Optimal data downloads that im-

prove effective throughputs are achieved using up-

per levels of pzstd/zstd on connections with hig-

her network throughputs, and using upper levels of

lbzip2/pbzip2 and pxz/xz on connections with lower

network throughputs. The cost-effective data down-

loads are achieved using lbzip2/pbzip2 and pixz/pxz,

as well as using bzip2 and xz in case of low network

throughput.

The rest of the paper is organized as fol-

lows. Section 2 presents background and motivation.

Section 3 describes experimental evaluation. Section

4 presents the results of the evaluation. Section 5 sur-

veys related work. Finally, Section 6 draws conclusi-

ons and summarizes our ﬁndings.

2 BACKGROUND

2.1 Transfers in the Cloud Computing

Figure 1 illustrates main steps in ﬁle transfers initiated

from a workstation. A data ﬁle can be transferred to

or from the cloud uncompressed or compressed. For

PEC 2018 - International Conference on Pervasive and Embedded Computing

uncompressed transfers, an uncompressed ﬁle (UF)

is uploaded or downloaded over a network directly.

For compressed uploads, the uncompressed ﬁle is ﬁrst

compressed locally, and then a compressed ﬁle (CF)

is uploaded to the cloud over the network. For com-

pressed downloads, a compressed version of the re-

quested ﬁle is downloaded from the cloud, and then

the compressed ﬁle is decompressed locally. When a

compressed version of the requested ﬁle is not availa-

ble in the cloud, the ﬁle is compressed in the cloud,

and then the compressed ﬁle is downloaded and de-

compressed on the workstation. The ﬁle transfer and

local (de)compression tasks on the workstation and

the cloud are often overlapped. Compressed transfers

utilize one of available compression utilities. Each

compression utility supports a range of compression

levels that allow us to trade off speed for compression

ratio: lower levels favor speed, whereas higher levels

favor compression ratio.

We consider six common compression utilities, as

well as their parallel implementations, listed in Ta-

ble 1. We have selected relatively fast gzip, lzop, lz4,

and zstd utilities, as well as bzip2 and xz, which pro-

vide high compression ratios. As modern worksta-

tions include multicore processors, we also consider

pigz, pbzip2, lbzip2, pxz, pixz, and pzstd, which are

parallel versions of gzip, bzip2, xz, and zstd, respecti-

vely. For each utility, we consider at least three com-

pression levels: low (1), medium (6), and high (9, 12,

19).

To evaluate the performance of individual com-

pression (utility, level) pairs, we measure the total

time and the total costs of doing the ﬁle transfer

to and from the cloud. The total time, in general,

includes the following components: (i) workstation

overhead time, (ii) network connection setup time,

(iii) ﬁle transmission time, and (iv) cloud overhead

time. Instead of reporting execution time, the ef-

fective throughputs, Th.CUP [Th.CDW], are deﬁned

as the ratio between the uncompressed ﬁle size in me-

gabytes and the total time needed to complete the ﬁle

Table 1: Lossless compression utilities.

Utility Levels Version Notes

gzip 1-9 (6) 1.6 DEFLATE (Ziv-Lempel, Huffman)

lzop 1-9 (6) 1.03 LZO (Lempel-Ziv-Oberhumer)

lz4 1-9 (6) r128 LZ4

bzip2 1-9 (6) 1.0.6 RLE+BWT+MTF+Huffman

xz 0-9 (6) 5.1.0 LZMA2

zstd 1-19 (3) 1.1.4 Huff0+FSE+LZ77

pigz 1-9 (6) 2.3.3 Parallel implementation of gzip

pbzip2 1-9 (9) 1.1.12 Parallel implementation of bzip2

lbzip2 1-9 (9) 1.1.12 Parallel implementation of bzip2

pxz 0-9 (9) 5.1.0 Parallel implementation of xz

pixz 0-9 (9) 1.0.6 Parallel implementation of xz

pzstd 1-19 (3) 1.1.4 Parallel implementation of zstd

transfer. The total costs depend on the utilization fees

and transfer fees charged by the cloud provider and

heavily depend on the effective throughput and achie-

vable compression ratio.

2.2 AWS Cloud Computing Platform

To facilitate cloud computing, we use Amazons Elas-

tic Cloud Computing (EC2) that provides computati-

onal and storage resources across a number of global

locations. Cloud instances in North Virginia and To-

kyo (Table 2) are created using a compute and me-

mory optimized m4.xlarge Linux instance type. The

selected locations of the clouds have a direct impact

on the network throughput and time to set up the net-

work connection from the local workstation. For the

geographically close instance in North Virginia, the

observed network throughput is 18.4 MB/s for uplo-

ads and 80.6 MB/s for downloads, and the time to set

up a secure connection (ssh) is less than 1 s. For

the cloud instance in Tokyo, the observed network

throughput is 8.4 MB/s for uploads and 10.5 MB/s for

downloads, and the time to set up a connection is 5 s.

In both cases, the network throughputs can be further

capped by a high-intensity network trafﬁc and a low

network priority.

Per AWS EC2 on-demand pricing model, the user

is charged the utilization fee (per hour) based on se-

lected instance type and location, and the download-

out fee charged per GB of data transfer. For exam-

ple, the utilization fees for the selected North Virgi-

nia and Tokyo clouds are $0.239 and $0.348 per hour,

respectively. The download-out fees are $0.09 and

$0.14 per GB, respectively. File transfers to and bet-

ween the cloud instances in the same region are free.

The use of the cloud can be divided into two

categories, a service-based and compute-based on-

demand usage. A service-based usage refers to al-

ways on cloud instance, with primary examples such

as web and ﬁle servers. In this case, the utilization

cost becomes an invariable constant. A compute-

based usage refers to on-demand cloud instance, swit-

ched on only for servicing computational tasks sub-

mitted by the workstation. In this case, the total uti-

lization cost can be optimized by using AWS scripts

to start an instance, upload inputs, execute job on the

cloud, download results, and ﬁnally shutdown the in-

stance.

Table 2: AWS EC2 Cloud Instances.

Instance

Location

Distance Utilization &

Type (miles) Download Cost

NV m4.xlarge North Virginia 600 $0.239 / $0.09

TK m4.xlarge Tokyo 7,000 $0.348 / $0.14

On Effectiveness of Compressed File Transfers to/from the Cloud: An Experimental Evaluation

With on-demand cloud usage, the total cost of

uploading a ﬁle to the cloud (TC.UUP [TC.CUP])

depends on the ﬁle size, US, utilization fee, Util

FEE

(expressed in $/s), and the effective upload throug-

hput, as shown in (1). Because AWS does not charge

for data uploads, the optimal costs are achieved by

transfer modes with the highest upload throughputs.

The total costs of downloading a ﬁle from the cloud

(TC.UDW [TC.CDW]) depend on the ﬁle size, utili-

zation fee, download-out fee, Dout

FEE

(expressed in

$/MB), effective download throughput, and the com-

pression ratio, CR, as shown in (2). To compare efﬁ-

ciency of cloud utilization by each (utility, level) pair,

we also deﬁned cost efﬁciency for upload (CE.UUP

[CE.CUP]), as shown in (3), and cost efﬁciency for

download (CE.UDW [CE.CDW]), as shown in (4).

Cost efﬁciency is expressed in megabytes per dollar

(MB/$).

TC.UUP[TC.CUP] = U S · (

Util

FEE

/3600

T h.UU P[T h.CUP]

)

(1)

TC.U DW [TC.CDW ] = US · (

Util

FEE

/3600

1 + T h.DW ·

T.SC

Dout

FEE

/1024

)

(2)

CE.UUP[CE.CUP] =

TC.UUP[TC.CUP]

(3)

CE.U DW [CE.CDW ] =

TC.U DW [TC.CDW ]

(4)

3 EXPERIMENTAL EVALUATION

Experimental evaluation involves performing com-

pressed ﬁle transfers initiated from a workstation to

and from the cloud instances in North Virginia (NV)

and Tokyo (TK). The m4.xlarge cloud instance type

used in both cases features 2.3 GHz Intel Xeon E5-

2686 with 4 virtual cores, 16 GB of RAM memory

and 750 Mbps of dedicated bandwidth. The worksta-

tion initiating transfers is a Dell PowerEdge T110 II

featuring quad-core Intel Xeon E3-1240 v2 and 8 GB

of RAM memory.

Each transfer mode is augmented to measure the

total transfer times and thus the effective throug-

hputs. The total costs for using the cloud are cal-

culated using measured effective throughputs, local

compression ratios of each (utility, level) pair, and

cloud fees. To measure effectiveness of each utility

in terms of speedup, the effective throughputs achie-

ved by the (utility, level) pair, Th.CUP [Th.CDW],

are compared to the effective throughputs of uncom-

pressed transfers, Th.UUP [Th.UDW], and the default

compressed transfers with gzip -6, Th.CUP(gzip -6)

[Th.CDW(gzip -6)]. To measure cost savings, the per-

centage of saved cost is calculated relative to the cost

of doing uncompressed transfers as shown in (3) and

relative to the cost of doing default compression trans-

fers as shown in (4). To capture effect of network,

the experiments are repeated using uncapped, 5 MB/s,

and 2 MB/s LAN network connections.

TC.UUP − TC.CUP]

TC.UUP

[

TC.U DW − TC.CUP

TC.U DW

]

(5)

TC.CUP(gzip − 6) − TC.CUP]

TC.CUP(gzip − 6)

[

TC.CDW (gzip − 6) − TC.CUP

TC.CDW (gzip − 6)

]

(6)

Table 3: Datasets to characterize local compression.

ID Dataset Type

Size

Notes

D0 netcdf bin 7.2 Relief data in NetCDF format

D1 seq.all txt 6.7 DNA sequence data in text

D2 wikipages xml 8.1 Wikipedia archived pages

Datasets. A diverse collection of ﬁles representa-

tive of data transfers to and from the cloud is compiled

as shown in Table 3. The datasets consist of text ﬁ-

les containing DNA sequence data from the UniGene

(NCBI, 2017) project, binary ﬁles containing Earths

surface relief data collected by the National Oceanic

and Atmospheric Administration (NOAA) (Amante,

C. and B. W. Eakins, 2009), and XML ﬁles contai-

ning web data with human-readable English text col-

lected from archived pages on Wikipedia (Wikipedia,

2017). Each dataset includes 20 ﬁles varying in size.

Files range in size from 1.61 MB to 1.87 GB with the

average and median being 592.36 MB and 333.29 MB

for wikipages ﬁles, 263.77 MB and 157.38 MB for

netcdf ﬁles, and 216.5 MB and 121.76 MB for seq.all

ﬁles.

4 RESULTS

4.1 Local Compression Ratios and

Throughputs

Table 4 shows the compression ratios for all consi-

dered sequential compression utilities and levels for

the input datasets. The compression ratios are avera-

ged from 20 points to 2 points that encompass ﬁles

smaller than 100 MB and larger than 100 MB in size.

The parallel utilities achieve identical or very similar

compression ratios as their sequential counterparts.

lzop has the lowest compression ratios, followed by

PEC 2018 - International Conference on Pervasive and Embedded Computing

Table 4: Compression ratios.

gzip lzop lz4 bzip2 xz zstd

ID -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -19

Less than 100 MB

D0 2.50 2.43 2.59 1.40 1.40 2.26 1.69 2.13 2.13 4.31 6.11 6.47 4.59 4.21 6.06 2.56 3.39 4.60

D1 2.25 2.17 2.30 1.30 1.30 2.01 1.53 1.90 1.90 3.80 5.27 5.57 3.97 3.66 5.16 2.26 2.95 3.94

D2 2.28 2.20 2.34 1.31 1.31 2.04 1.55 1.93 1.93 3.88 5.40 5.71 4.06 3.74 5.29 2.30 3.01 4.03

More than 100 MB

D0 2.40 2.32 2.49 1.41 1.42 2.25 1.61 2.09 2.09 3.83 4.77 4.93 3.95 3.55 4.59 2.38 2.91 3.88

D1 2.39 2.31 2.49 1.41 1.41 2.25 1.60 2.09 2.09 3.81 4.74 4.90 3.93 3.53 4.55 2.37 2.89 3.86

D2 2.42 2.40 2.58 1.44 1.45 2.32 1.64 2.15 2.16 3.86 4.79 4.94 3.97 3.62 4.67 2.45 2.98 3.96

Table 5: Local compression throughputs.

pigz lzop lz4 lbzip2 pixz pzstd

ID -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -19

Less than 100 MB

D0 188 45.7 17.9 148 142 3.22 128 31.9 30.0 71.4 81.7 78.8 43.3 3.60 1.10 468 100 4.10

D1 254 67.5 20.9 212 212 2.04 190 22.6 6.03 61.1 59.2 57.5 67.0 3.30 1.16 633 195 3.63

D2 209 85.9 2.09 113 111 6.86 110 26.5 21.1 55.4 52.0 48.6 42.0 3.76 1.20 399 96.4 3.70

More than 100 MB

D0 168 32.1 14.5 157 141 2.55 99.5 28.3 27.7 68.9 81.2 80.4 36.9 4.96 2.95 457 99.6 7.12

D1 262 72.5 26.0 211 211 2.57 188 24.0 7.10 64.3 63.7 62.3 61.7 5.92 3.04 928 254 7.75

D2 239 93.1 2.20 123 121 7.30 122 28.5 22.5 63.6 60.7 58.0 45.7 6.68 3.70 587 116 7.79

Table 6: Local decompression throughputs.

pigz lzop lz4 lbzip2 pixz pzstd

ID -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -9 -1 -6 -19

Less than 100 MB

D0 83.0 93.8 95.2 120 122 155 143 153 151 196 109 90.2 43.3 3.60 1.10 188 233 287

D1 95.2 137 135 150 150 239 198 235 269 189 118 107 45.2 68.7 85.6 424 429 463

D2 78.6 89.4 86.3 95.4 94.2 123 138 168 160 165 88.6 78.2 31.6 38.2 47.5 183 258 279

More than 100 MB

D0 76.6 91.0 99.6 122 123 169 138 183 182 221 98.5 80.5 36.9 4.96 2.95 193 243 300

D1 107 157 161 166 168 325 246 349 372 240 133 116 43.0 70.4 93.5 583 631 809

D2 101 108 109 121 119 157 188 244 243 204 113 96.7 33.9 40.1 50.1 267 332 393

lz4, zstd, and gzip. The highest compression ratios

are achieved by bzip2 and xz. For all utilities, hig-

her compression levels result in higher compression

ratios. Compression ratios range from 1 (lzop -1) to

10.12 (bzip2 -9) for netcdf, 2.26 (lzop -1) to 14.39 (xz

-9) for seq.all, and 1.19 (lzop -9) to 7.73 (xz -9) for

wikipages.

Table 5 and Table 6 show the local compression

and decompression throughputs. The throughputs of

the parallel utilities are shown in place of the sequen-

tial counterparts because they achieve higher throug-

hputs. Both lbzip2 and pixz outperform pbzip2 and

pxz through better parallel implementation. The hig-

hest local compression throughputs are achieved by

pzstd -1 reaching up to 746.05, 1557.58, and 1092.03

MB/s for netcdf, seq.all, and wikipages respectively.

The lowest compression throughputs are achieved by

lbzip2, pbzip2, pixz, pxz, and their sequential counter-

parts. The highest local decompression throughputs

are achieved by pzstd and zstd with high compression

levels, reaching up to 609.75, 1530.51, and 810.02

MB/s for netcdf, seq.all, and wikipages, respectively.

After comparing the compression ratios and ta-

king into account better compression and decompres-

sion throughputs, we ﬁnd that zstd and pzstd can ef-

fectively replace both gzip and pigz for all upload and

download transfers considered in this evaluation.

4.2 Compressed Upload/Download

Throughputs

Uploads. Figure 2 shows the effective throughput

for compressed uploads of all selected ﬁles (netcdf,

seq.all, wikipages) to the NV instance with the un-

capped network connection. The plots show that

the effective throughput saturates for the larger ﬁ-

les, approaching the product of the compression ratio,

CR, and the network connection upload throughput,

Th.UUP. For top performing utilities and levels, the

effective throughput increases with increase of input

ﬁle size, ranging from 5.02 to 171.43 MB/s (lbzip2,

zstd, pzstd). Availability of compute resources on the

workstation (quad-core processor and 8 GB of RAM

memory) allows parallel utilities to signiﬁcantly in-

crease the effective throughput of compressed uplo-

ads when using higher compression levels over their

On Effectiveness of Compressed File Transfers to/from the Cloud: An Experimental Evaluation

Figure 2: Effective throughput of compressed uploads (Uncapped, North Virginia).

Table 7: Maximum upload throughput speedups and optimal modes (North Virginia & Tokyo).

Uncapped 5 MB/s 2 MB/s

Clouds NV TK NV TK NV TK

TH.CUP/TH.UUP

D0 8.84 lbzip2 9.56 lbzip2 7.38 lbzip2 9.07 lbzip2/pbzip2 8.93 lbzip2 5.69 lbzip2/pbzip2

D1 5.28 pzstd 10.2 lbzip2/pbzip2 10.3 pzstd2/zstd 8.74 pzstd/zstd 10.5 zstd 9.92 pzstd/zstd

D2 9.44 pzstd/lbzip2 25.6 lbzip2 6.63 pbzip2/zstd 22.5 pzstd/zstd 5.92 pbzip2 5.16 pzstd/zstd

TH.CUP/TH.CUP(gzip -6)

D0 10.3 lbzip2 4.69 lbzip2 4.76 lbzip2 7.41 lbzip2/pbzip2 4.26 lbzip2 2.1 lbzip2/pbzip2

D1 7.73 pzstd 3.08 lbzip2/pbzip2 2.19 pzstd2/zstd 2.03 pzstd/zstd 1.35 zstd 1.33 pzstd/zstd

D2 8.84 pzstd/lbzip2 1.47 lbzip2 3.74 pbzip2/zstd 17.4 pzstd/zstd 2.11 pbzip2 1.39 pzstd/zstd

sequential counterparts. In some cases (e.g., lbzip2

and pzstd), higher compression levels outperform lo-

wer compression levels. For example, the maximum

throughput achieved by the best performing zstd pair

(-3) is 145.34 MB/s, while the maximum throughput

achieved by the best performing pzstd pair (-9) is

171.43 MB/s. This observation is deviation from a

similar study done on the mobile devices (Milenkovi

et al., 2013; Dzhagaryan et al., 2013; Dzhagaryan

et al., 2015), where lower levels consistently perfor-

med better on upload. For compressed uploads of all

selected ﬁles to the TK instance with the uncapped

network connection, the effective throughput ranges

from 0.80 to 58.60 MB/s, similarly approaching the

product of the compression ratio and the network con-

nection upload throughput.

For uploads to the NV instance with the uncapped

network connection, the optimal (utility, level) pairs

are lbzip2 -9 for netcdf, pzstd -9 for seq.all, and pzstd

-9 for small and lbzip2 -9 for large ﬁles in wikipa-

ges. High network throughput, low connection time,

and abundance of computer power in the workstation

favor utilities with solid throughputs and compres-

sion ratios. Files with higher compressibility (e.g.,

netcdf, and wikipages) and larger sizes favor lbzip2.

For transfers with the 5 MB/s connection, the optimal

pair for netcdf remains unchanged, whereas seq.all

and wikipages achieve high throughputs with zstd and

pbzip2 for larger ﬁles. For transfers with the 2 MB/s

connection, the optimal pairs are lbzip2 -9 for netcdf,

zstd -9 for seq.all, and pbzip2 -9 for wikipages.

For transfers to the TK instance that has lower net-

work throughput and a larger connection time than

the NV instance, the highest effective throughput is

achieved by utilities with high compression ratios

(e.g., pbzip2, lbzip2, pxz). For transfers with un-

capped network connection, the optimal modes are

lbzip2 -9 for netcdf, upper levels of pbzip2 and lbzip2

for seq.all, and lbzip2 -6 and -9 for wikipages. For

transfers with 5 MB/s an d 2 MB/s network, a slo-

wer pbzip2 -9 is selected more frequently over lbzip2

-9 for netcdf, while for seq.all and wikipages, optimal

throughputs are achieved with upper levels of zstd and

pzstd (-9, -12).

Table 7 shows optimal (utility, level) pairs and the

maximum throughput speedups relative to the uncom-

pressed and default compressed uploads to the NV

and TK instances. In both cases, speedups are greater

at higher network throughputs. Relative to uncom-

pressed uploads to the NV instance, the maximum

speedups range from 5.28 (netcdf ) to 9.44 (wikipa-

ges) on uncapped network, from 6.64 (wikipages) to

10.3 (seq.all) on 5 MB/s network, and from 5.92 (wi-

kipages) to 10.5 (seq.all) on 2 MB/s network. Rela-

PEC 2018 - International Conference on Pervasive and Embedded Computing

Figure 3: Effective throughput of compressed downloads (Uncapped, North Virginia).

Table 8: Maximum download throughput speedups and optimal modes (North Virginia & Tokyo).

Uncapped 5 MB/s 2 MB/s

Clouds NV TK NV TK NV TK

TH.CDW/TH.UDW

D0 3.82 pzstd 4.64 pbzip2 7.65 lbzip2 4.74 bzip2/pbzip2 8.98 lbzip2 6.14 bzip2/pbzip2

D1 7.87 pzstd 9.37 xz 14.01 xz 19.57 xz 14.23 xz 13.2 xz

D2 5.31 zstd/pzstd 5.76 xz 7.28 xz 6.26 xz 7.43 xz 6.48 xz

TH.CDW/TH.CDW(gzip -6)

D0 3.61 pzstd 6.78 pbzip2 2.68 lbzip2 4.43 bzip2/pbzip2 2.85 lbzip2 13.0 bzip2/pbzip2

D1 2.94 pzstd 4.66 xz 1.8 xz 13.8 xz 1.82 xz 2.52 xz

D2 2.66 zstd/pzstd 2.91 xz 1.84 xz 2.27 xz 1.87 xz 5.6 xz

tive to uncompressed uploads to the TK instance, the

maximum speedups range from 9.56 (netcdf ) to 25.6

(wikipages), from 8.74 (seq.all) to 22.5 (wikipages),

and from 5.16 (wikipages) to 9.92 (seq.all) on uncap-

ped, 5 MB/s, and 2 MB/s network respectively. Rela-

tive to default compressed uploads to the NV instance,

the maximum speedups range from 8.84 to 10.3 (un-

capped), from 2.19 to 4.76 (5 MB/s), and from 2.11

to 4.26 (2 MB/s) respectively. Relative to the default

compressed uploads to the TK instance, the maximum

speedups range from 1.47 to 4.9 (uncapped), 2.03 to

17.4 (5 MB/s), and from 1.33 to 2.1 (2 MB/s).

Downloads. Figure 3 shows the effective throug-

hput for compressed downloads of all selected ﬁles

(netcdf, seq.all, wikipages) from the NV instance with

the uncapped network connection. Due to high local

decompression throughputs, the effective throughputs

reach higher levels than for uploads and saturate si-

milarly across most compression levels within each

utility. For top performing utilities and levels, the ef-

fective throughput increases with increase of input ﬁle

size, ranging from 5.66 MB/s to 579.58 MB/s (zstd,

pzstd). For compressed downloads of all selected ﬁ-

les to the TK instance with the uncapped network

connection, the effective throughput ranges from 0.86

to 93.54 MB/s. Limiting the network throughput on

both NV and TK instance similarly limits maximum

throughput to the product of the compression ratio

and the network connection download throughput.

For downloads from the NV instance with uncap-

ped network connection, the optimal (utility, level)

pairs are pzstd -19 for netcdf, pzstd -19 for seq.all,

and pzstd -19 for small and zstd -19 for large ﬁles in

wikipages. With lower network throughputs (5 MB/s

and 2 MB/s), favor shifts toward slower or sequential

(utility, level) pairs with stronger compression. For

downloads with 5 MB/s and 2 MB/s network, the op-

timal pair is lbzip2 -9 for netcdf, and xz -9 for seq.all

and wikipages. Similarly, for downloads from Tokyo

instance slower or sequential (utility, level) pairs with

stronger compression are selected (pbzip2, bzip2, xz).

For downloads with uncapped network connection,

the optimal modes are pbzip2 -9 for netcdf, xz -9 for

seq.all and wikipages. For downloads with 5 MB/s

and 2 MB/s networks, pbzip2 -9 for smaller and bzip2

-9 for larger ﬁles are selected for netcdf, whereas the

optimal pairs for seq.all and wikipages remain un-

changed.

Table 8 shows the maximum throughput speedups

and optimal (utility, level) pairs when being compa-

red to throughputs achieved with uncompressed and

default compressed downloads from the NV and TK

instances. In both cases, speedups are greater at the

lower network throughput. Relative to uncompres-

On Effectiveness of Compressed File Transfers to/from the Cloud: An Experimental Evaluation

Figure 4: Total cloud cost of compressed uploads (Uncapped, North Virginia).

Table 9: Maximum upload cost savings and optimal modes (North Virginia & Tokyo).

Uncapped 5 MB/s 2 MB/s

Clouds NV TK NV TK NV TK

(%) Relative to TC.UUP

D0 99.2 lbzip2 86.6 lbzip2 89.6 lbzip2 86.9 lbzip2/pbzip2 89.8 lbzip2 87.9 lbzip2/pbzip2

D1 99.6 pzstd 89.9 lbzip2/pbzip2 90.5 pzstd2/zstd 90.3 pzstd/zstd 90.5 zstd 90.4 pzstd/zstd

D2 99.2 pzstd/lbzip2 85.1 lbzip2 83.7 pbzip2/zstd 89.3 pzstd/zstd 83.3 pbzip2 83.0 pzstd/zstd

Relative to TC.CUP(gzip -6)

D0 90.3 lbzip2 65.2 lbzip2 67.8 lbzip2 65.5 lbzip2/pbzip2 67.1 lbzip2 62.9 lbzip2/pbzip2

D1 87.1 pzstd 31.0 lbzip2/pbzip2 30.3 pzstd2/zstd 29.5 pzstd/zstd 25.8 zstd 25.4 pzstd/zstd

D2 88.7 pzstd/lbzip2 71.2 lbzip2 48.2 pbzip2/zstd 78.9 pzstd/zstd 40.3 pbzip2 33.1 pzstd/zstd

sed downloads from the NV instance, the maximum

speedups range from 3.82 (netcdf ) to 5.31 (wikipages)

on uncapped network, from 7.28 (wikipages) to 14.01

(seq.all) on 5 MB/s network, and from 7.43 (wikipa-

ges) to 14.23 (seq.all) on 2 MB/s network. Relative to

uncompressed downloads from the TK instance, the

maximum speedups range from 4.64 (netcdf ) to 9.37

(seq.all), from 4.74 (netcdf ) to 19.57 (seq.all), and

from 6.14 (netcdf ) to 13.2 (seq.all) on uncapped, 5

MB/s, and 2 MB/s network. Relative to default com-

pressed downloads from the NV instance, the maxi-

mum speedups range from 2.66 to 3.61 (uncapped),

from 1.8 to 2.68 (5 MB/s), and from 1.82 to 2.85

(2 MB/s) respectively. Relative to default compres-

sed downloads from the TK instance, the maximum

speedups range from 2.97 to 6.78 (uncapped), 2.27 to

13.8 (5 MB/s), and from 2.52 to 13.0 (2 MB/s).

4.3 Compressed Upload/Download

Costs

Uploads. Figure 4 shows the total cloud cost for com-

pressed uploads of all selected ﬁles (netcdf, seq.all,

wikipages) to the NV instance with the uncapped net-

work connection. The plots show that the total cloud

cost increases with increase of input ﬁle size, ranging

from $0.0005 to $6.7119. With the cloud cost depen-

ding only on the utilization fees (Util

FEE

), the derived

cost efﬁciency has similar amount of saturation and

ranking among different (utility, level) pairs as those

of the effective throughput. For top performing utili-

ties and levels, the cloud cost efﬁciency ranges from

93.5 to 25,821 MB/$ (lbzip2, zstd, pzstd). For com-

pressed uploads of all selected ﬁles to the TK instance

with the uncapped network connection, the total cloud

cost ranges from $0.002 to $9.80 and the cloud cost

efﬁciency ranges from 35.08 to 6061.95 MB/$. Limi-

ting the network throughput on both NV and TK in-

stance similarly increases total cost and reduces cost

efﬁciency.

Table 9 shows the maximum cost savings when

being compared to the total costs of uncompressed

and to the default compressed uploads to the NV and

TK instances. With no charge for uploading ﬁles to

AWS instances, the (utility, level) pairs that achieve

the highest throughputs achieve the highest cost sa-

vings. The maximum cost savings relative to the cost

of doing uncompressed ﬁle uploads to the NV in-

stance range from 83.3 - 90.5% (2 MB/s LAN) to

99.2 - 99.6% (uncapped). The maximum cost savings

achieved with optimal compressed uploads to the TK

instance are similar to those of the NV instance for

PEC 2018 - International Conference on Pervasive and Embedded Computing

Figure 5: Total cloud cost of compressed downloads (Uncapped, North Virginia).

Table 10: Maximum download cost savings and optimal modes (North Virginia & Tokyo).

Uncapped 5 MB/s 2 MB/s

Clouds NV TK NV TK NV TK

(%) Relative to TC.UDW

D0 89.1 lbzip2 86.8 pbzip2 89.7 lbzip2 86.8 pbzip2 89.8 lbzip2 88.2 pbzip2

D1 92.6 xz 92.8 xz 93.0 xz 92.8 xz 93.0 xz 92.9 xz

D2 86.3 xz 86.1 xz 86.9 xz 86.1 xz 86.9 xz 86.1 xz

Relative to TC.CDW(gzip -6)

D0 65.6 lbzip2 68.5 pbzip2 66.9 lbzip2 68.5 pbzip2 67.1 lbzip2 83.7 pbzip2

D1 42.9 xz 50.9 xz 45.2 xz 50.9 xz 45.2 xz 67.7 xz

D2 45.7 xz 48.6 xz 47.4 xz 48.6 xz 47.5 xz 75.9 xz

2 MB/s and 5 MB/s network throughputs, but are lo-

wer with the uncapped network due to lower network

throughput (85.1 - 89.9%).

The maximum cost savings relative to the cost of

doing the default compressed uploads to the NV in-

stance range from 25.8 - 65.5% (2 MB/s LAN) to 88.7

- 90.7% (uncapped). The maximum cost savings re-

lative to the cost of the default compressed uploads to

the Tokyo cloud are lower, ranging from 25.4 - 62.9%

(2 MB/s LAN) to 31.0 - 71.2% (uncapped).

Downloads. Figure 5 shows the total cloud cost

for compressed downloads of all selected ﬁles (netcdf,

seq.all, wikipages) from the TK instance with the un-

capped network connection. The total cloud cost ran-

ges from $0.0089 to $8.5198. For top performing uti-

lities and levels, the derived cloud cost efﬁciency ran-

ges from 107.25 to 1522.78 MB/$ (lbzip2, pxz). For

compressed downloads of all selected ﬁles to the TK

instance with the uncapped network connection, the

total cloud cost ranges from $0.0387 to $14.1553 and

the cloud cost efﬁciency ranges from 49.99 to 949.18

MB/$. Limiting the network throughput on both NV

and TK instance similarly increases total cost and re-

duces cost efﬁciency.

Table 10 shows the maximum cost savings when

being compared to the total costs of uncompressed

and default compression downloads from the NV and

TK instances. The most cost-effective (utility, level)

pairs are the ones that achieve a balance between good

compression ratio and high throughput. For down-

loads from the NV instance, the most cost-effective

(utility, level) pairs are lbzip2 -9 for netcdf ﬁles, and

xz -9 for seq.all and wikipages ﬁles. For downloads

from the TK instance the most cost-effective (utility,

level) pairs are pbzip2 -9 for netcdf ﬁles, and xz -9

for seq.all and wikipages ﬁles, regardless of network

throughput.

The maximum cost savings relative to the cost of

doing uncompressed ﬁle downloads from the NV in-

stance range from 86.9 - 93.0% (2 MB/s LAN) to 86.3

- 92.6% (uncapped). Similar cost savings are achie-

ved by the most cost-effective compressed downloads

from the TK instance. The maximum cost savings re-

lative to the cost of doing default compressed down-

loads are greater at the lower network throughput in

both cases. The maximum cost savings relative to the

default compressed downloads from the NV instance

range from 42.9 - 65.6% (uncapped) to 45.2 - 67.1%

(2 MB/s LAN). With lower network throughputs and

longer network setup times, higher cost savings are

achieved by the most cost-effective compressed do-

wnloads from the TK instance, ranging from 48.6 -

68.5% (uncapped) to 67.7 - 83.7% (2 MB/s LAN).

Cloud cost for download depends on both utiliza-

On Effectiveness of Compressed File Transfers to/from the Cloud: An Experimental Evaluation

tion fee and download-out fee. Due to fast decom-

pression and Due to additional download-out cost, to-

tal cloud cost of download increases much more ex-

ponentially with increasing ﬁle size.

5 RELATED WORK

We are aware of several studies that investigate use

of lossless data compression and decompression in

improving the overall effectiveness of ﬁle transfers

between edge devices and the cloud. This includes

measurement-based studies that evaluated common

and emerging general-purpose compression utilities,

and studies that introduced either analytical models or

run-time techniques for deciding whether to use com-

pressed transfer or not. Studies have been conducted

on workstations, as in this paper, and on embedded

and mobile devices.

The ﬁrst set of studies, most closely related to

own work, focused on optimizing data transfers to

and from the cloud while being initiated from a work-

station. Optimizing data transfers calls for increasing

communication throughput, reducing connection la-

tency, making effective use of cloud storage, and re-

ducing cloud costs (Abadi, 2009; Abu-Libdeh et al.,

2010; Sboner et al., 2011). Measurement based study

performed by Nicolae (Nicolae, 2010) compares un-

compressed transfers with lzo and bzip2 compressed

transfers and proposes precompression of header to

select either uncompressed or optimally compressed

transfer. Work by Bonﬁeld and Mahoney (Bonﬁeld

and Mahoney, 2013) compares several general pur-

pose compression (gzip, bzip2) and genome speciﬁc

compression utilities using DNA sequencing data. Bi-

cer and others (Bicer et al., 2013) develop new com-

pression utility and compare it to common compres-

sion utilities (gzip, bzip2, LZO, LZMA) for upload and

download of scientiﬁc data. Over all, measurement-

based studies showed that lossless compressed trans-

fers can increase throughput and reduce overall cloud

costs during transfers. Similar to our work, datasets

with larger representative ﬁles are considered for in-

vestigations. Compared to their work, we have cove-

red a wider range of sequential and parallel lossless

compression utilities, including pzstd/zstd developed

speciﬁcally for use in data centers (Facebook, 2017).

Our experiments were conducted on two AWS EC2

Cloud instances in geographically different locations,

allowing us to extract not only effective throughput,

but also effective cost for a range of input ﬁles and

network conditions.

The second set of studies focused on optimizing

data transfers initiated from battery-powered embed-

ded and mobile devices. There, energy-efﬁciency be-

comes a priority followed by similar goals of incre-

asing communication throughput and reducing con-

nection latency. The datasets in those studies are se-

lected to better represent data transfers initiated to

and from mobile devices, which included raw ima-

ges, physiological data (csv), code, binary, and text.

Study conducted almost a decade ago by Barr and

Asanovi

c (Barr and Asanovi

c, 2003; Barr and Asa-

novi

c, 2006) investigated energy efﬁciency of los-

sless data compression on a wireless handheld devi-

ces. Several studies introduced run-time techniques

for deciding whether to use compressed transfer or

not (Krintz and Sucu, 2006; Nicolae, 2010; Harnik

et al., 2013). Study by Krintz and others (Krintz and

Sucu, 2006) introduced technique that analyzes strea-

ming data in-transit and makes decisions about use of

compression. Works by Nicolae and Harnik (Nicolae,

2010; Harnik et al., 2013) propose techniques which

rely on data pre-compression to estimated compres-

sion ratios and use it as the deciding factor in selecting

optimal transfer modes.

Our past work on embedded and modern mobile

devices performed extensive measurement-based eva-

luation of common general-purpose compression uti-

lities (Milenkovi

c et al., 2013; Dzhagaryan et al.,

2013; Dzhagaryan and Milenkovi

c, 2015; Dzhaga-

ryan et al., 2015), introduced analytical models des-

cribing effective throughput and energy efﬁciency for

uncompressed as well as for compressed ﬁle transfers

(Dzhagaryan and Milenkovi

c, 2016), and introduced

a framework for selecting an optimal transfer mode

(Dzhagaryan and Milenkovi

c, 2017). Those studies

have also been supported by development of environ-

ment and tools for automated measurement of energy

consumption on mobile and embedded devices (Mi-

losevic et al., 2013; Dzhagaryan et al., 2016a; Dzha-

garyan et al., 2016b).

6 CONCLUSIONS

This paper presents an experimental evaluation of

general-purpose compression utilities for a range of

data ﬁles transferred between an edge workstation and

the cloud. The goal of this evaluation is to quantify ef-

fectiveness of individual compression utilities and to

provide guidelines for achieving the optimal throug-

hput or the lowest costs. The results of our study de-

monstrate that reliance on a single (utility, level) pair

is not an answer for achieving optimal throughput or

costs when transferring ﬁles to or from the cloud.

Selecting the optimal compression (utility, level)

pair for speciﬁc ﬁle size and type, network con-

PEC 2018 - International Conference on Pervasive and Embedded Computing

nection, cloud fees, and workstation performance can

signiﬁcantly improve the effective throughput and

signiﬁcantly reduce cloud cost when compared to un-

compressed and the default compressed ﬁle trans-

fers. Throughput-effective upload and download mo-

des favor utilities that offer tradeoffs between com-

pression ratio and throughput, such as pzstd/zstd and

lbzip2/pbzip2, whereas cost-effective download mo-

des favor utilities with higher compression ratio, such

as lbzip2/pbzip2 and pxz/xz. These ﬁndings may

guide throughput and cost optimizations of big data

transfers in the cloud and encourage the development

of data transfer frameworks conscientious of the ex-

isting parameters for real-time selection of optimal

transfer modes.

ACKNOWLEDGEMENTS

This work has been supported by AWS Cloud Credits

for Research and in part by National Science Founda-

tion grants CNS-1205439 and CNS-1217470.

REFERENCES

Abadi, D. J. (2009). Data management in the cloud: li-

mitations and opportunities. IEEE Data Eng. Bull.,

32(1):3–12.

Abu-Libdeh, H., Princehouse, L., and Weatherspoon, H.

(2010). RACS: A case for cloud storage diversity.

In Proceedings of the 1st ACM Symposium on Cloud

Computing, SoCC ’10, pages 229–240. ACM.

Amante, C. and B. W. Eakins (2009). ETOPO1 1 arc-

minute global relief model: Procedures, data sources

and analysis.

Amazon (2017). Amazon web services (AWS) - cloud com-

puting services. https://aws.amazon.com/.

Apple (2017). Data compression | apple developer docu-

mentation. https://tinyurl.com/ybwc77su.

Barr, K. and Asanovi

c, K. (2003). Energy aware lossless

data compression. In Proceedings of the 1st Internati-

onal Conference on Mobile Systems, Applications and

Services (MobiSys’03), pages 231–244. ACM Press.

Barr, K. C. and Asanovi

c, K. (2006). Energy-aware lossless

data compression. ACM Transactions on Computer

Systems, 24(3):250–291.

Bicer, T., Yin, J., Chiu, D., Agrawal, G., and Schuchardt, K.

(2013). Integrating online compression to accelerate

large-scale data analytics applications. In 2013 IEEE

27th International Symposium on Parallel Distributed

Processing (IPDPS), pages 1205–1216.

Bonﬁeld, J. K. and Mahoney, M. V. (2013). Compression

of FASTQ and SAM format sequencing data. PLoS

ONE, 8(3):1–10.

Chen, M., Mao, S., and Liu, Y. (2014). Big data: A survey.

Mobile Networks and Applications, 19(2):171–209.

Chen, Y., Ganapathi, A., and Katz, R. H. (2010). To com-

press or not to compress - compute vs. IO tradeoffs

for mapreduce energy efﬁciency. In Proceedings of

the First ACM SIGCOMM Workshop on Green Net-

working, Green Networking ’10, pages 23–28. ACM.

Cyan (2013). RealTime data compression: Finite

state entropy - a new breed of entropy coder.

https://tinyurl.com/p5ehc54.

Dzhagaryan, A. and Milenkovi

c, A. (2015). On effecti-

veness of lossless compression in transferring mhe-

alth data ﬁles. In 2015 17th International Conference

on E-health Networking, Application Services (Healt-

hCom), pages 665–668.

Dzhagaryan, A. and Milenkovi

c, A. (2016). Models for

evaluating effective throughputs for ﬁle transfers in

mobile computing. In 2016 25th International Con-

ference on Computer Communication and Networks

(ICCCN), pages 1–9.

Dzhagaryan, A. and Milenkovi

c, A. (2017). A framework

for optimizing ﬁle transfers between mobile devices

and the cloud. In 2017 IEEE 28th Annual Internatio-

nal Symposium on Personal, Indoor, and Mobile Ra-

dio Communications (PIMRC), pages 1–7.

Dzhagaryan, A., Milenkovi

c, A., and Burtscher, M. (2013).

Energy efﬁciency of lossless data compression on a

mobile device: An experimental evaluation. In 2013

IEEE International Symposium on Performance Ana-

lysis of Systems and Software (ISPASS), pages 126–

127.

Dzhagaryan, A., Milenkovi

c, A., and Burtscher, M. (2015).

Quantifying beneﬁts of lossless compression utilities

on modern smartphones. In 2015 24th International

Conference on Computer Communication and Net-

works (ICCCN), pages 1–9.

Dzhagaryan, A., Milenkovi

c, A., Milosevic, M., and Jova-

nov, E. (2016a). An environment for automated mea-

surement of energy consumed by mobile and embed-

ded computing devices. Measurement, 94:103–118.

Dzhagaryan, A., Milenkovi

c, A., Milosevic, M., and Jova-

nov, E. (2016b). An environment for automated me-

asuring of energy consumed by android mobile de-

vices. In Proceedings of the 6th International Joint

Conference on Pervasive and Embedded Computing

and Communication Systems - Volume 1: PEC,, pages

28–39.

Facebook (2017). Zstandard - real-time data compression

algorithm. https://tinyurl.com/zf4vfz6.

Google (2017a). Google cloud computing, hosting services

& APIs. https://cloud.google.com/.

Google (2017b). google/brotli. https://github.com/google/

brotli.

Harnik, D., Kat, R., Margalit, O., Sotnikov, D., and Trae-

ger, A. (2013). To zip or not to zip: Effective resource

usage for real-time compression. In Proceedings of

the 11th USENIX Conference on File and Storage

Technologies, FAST’13, pages 229–241. USENIX.

Krintz, C. and Sucu, S. (2006). Adaptive on-the-ﬂy com-

On Effectiveness of Compressed File Transfers to/from the Cloud: An Experimental Evaluation

pression. IEEE Transactions on Parallel and Distri-

buted Systems, 17(1):15–24.

Microsoft (2017). Microsoft azure: Cloud computing plat-

form & services. https://tinyurl.com/nmxlnae.

Milenkovi

c, A., Dzhagaryan, A., and Burtscher, M. (2013).

Performance and energy consumption of lossless

compression/decompression utilities on mobile com-

puting platforms. In 2013 IEEE 21st Internatio-

nal Symposium on Modeling, Analysis Simulation

of Computer and Telecommunication Systems (MAS-

COTS), pages 254–263.

Milosevic, M., Dzhagaryan, A., Jovanov, E., and Milen-

kovi

c, A. (2013). An environment for automated po-

wer measurements on mobile computing platforms. In

Proceedings of the 51st ACM Southeast Conference,

ACMSE ’13, page 6. ACM.

NCBI (2017). UniGene - NCBI. http://tinyurl.com/

yjnsmok.

Nicolae, B. (2010). High throughput data-compression for

cloud storage. In Hameurlain, A., Morvan, F., and

Tjoa, A. M., editors, Data Management in Grid and

Peer-to-Peer Systems, number 6265 in Lecture No-

tes in Computer Science, pages 1–12. Springer Berlin

Heidelberg.

Sboner, A., Mu, X. J., Greenbaum, D., Auerbach, R. K., and

Gerstein, M. B. (2011). The real cost of sequencing:

higher than you think! Genome Biology, 12:125.

Wikipedia (2017). enwiki dump. http://tinyurl.com/zqffp3l.

PEC 2018 - International Conference on Pervasive and Embedded Computing