Professional Documents
Culture Documents
Textbook Advanced Computer Architecture 12Th Conference Aca 2018 Yingkou China August 10 11 2018 Proceedings Chao Li Ebook All Chapter PDF
Textbook Advanced Computer Architecture 12Th Conference Aca 2018 Yingkou China August 10 11 2018 Proceedings Chao Li Ebook All Chapter PDF
https://textbookfull.com/product/advanced-computer-
architecture-13th-conference-aca-2020-kunming-china-
august-13-15-2020-proceedings-dezun-dong/
https://textbookfull.com/product/advanced-computer-
architecture-11th-conference-aca-2016-weihai-china-
august-22-23-2016-proceedings-1st-edition-junjie-wu/
https://textbookfull.com/product/network-and-system-
security-12th-international-conference-nss-2018-hong-kong-china-
august-27-29-2018-proceedings-man-ho-au/
https://textbookfull.com/product/software-architecture-12th-
european-conference-on-software-architecture-ecsa-2018-madrid-
spain-september-24-28-2018-proceedings-carlos-e-cuesta/
Frontiers in Cyber Security First International
Conference FCS 2018 Chengdu China November 5 7 2018
Proceedings Fagen Li
https://textbookfull.com/product/frontiers-in-cyber-security-
first-international-conference-fcs-2018-chengdu-china-
november-5-7-2018-proceedings-fagen-li/
https://textbookfull.com/product/computational-science-
iccs-2018-18th-international-conference-wuxi-china-
june-11-13-2018-proceedings-part-ii-yong-shi/
https://textbookfull.com/product/computational-science-
iccs-2018-18th-international-conference-wuxi-china-
june-11-13-2018-proceedings-part-iii-yong-shi/
https://textbookfull.com/product/computer-engineering-and-
technology-22nd-ccf-conference-nccet-2018-yinchuan-china-
august-15-17-2018-revised-selected-papers-weixia-xu/
Advanced
Computer Architecture
12th Conference, ACA 2018
Yingkou, China, August 10–11, 2018
Proceedings
123
Communications
in Computer and Information Science 908
Commenced Publication in 2007
Founding and Former Series Editors:
Phoebe Chen, Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu,
Dominik Ślęzak, and Xiaokang Yang
Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
Junsong Yuan
University at Buffalo, The State University of New York, Buffalo, USA
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at http://www.springer.com/series/7899
Chao Li Junjie Wu (Eds.)
•
Advanced
Computer Architecture
12th Conference, ACA 2018
Yingkou, China, August 10–11, 2018
Proceedings
123
Editors
Chao Li Junjie Wu
Shanghai Jiao Tong University National University of Defense Technology
Shanghai Changsha
China China
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
It is a great pleasure and honor to present the proceedings of ACA 2018, the 12th
Conference on Advanced Computer Architecture. ACA is sponsored by the China
Computer Federation (CCF) and it is the flagship conference of the CCF Technical
Committee on Computer Architecture (TCArch). It has been one of the most important
academic conferences in the field of computer architecture in China since 1995.
The 2018 edition of ACA was held in the scenic area of Yingkou, a port city of the
Bohai Sea. The theme this year was “Intelligent Architecture: From the Cloud to the
Edge.” ACA 2018 created a forum for academic researchers and industry practitioners
in China to share their insights on the next-generation computing systems. We con-
tinued the trend of making ACA an inclusive and interactive event that features invited
keynotes, top paper presentation, poster showcase, and design competition, etc.
This year, we received over 120 paper registrations. Finally, there were 80 suc-
cessful submissions. Each submission was reviewed by three Program Committee
(PC) members on average. In all, 13 papers were rejected immediately in the first round
of review and 67 papers were sent out for a second round of review. Only the papers
with an average score of 3 (borderline) were considered for final inclusion, and
almost all accepted papers had positive reviews or at least one review with a score of 5
(accept) or higher. Finally, the PC decided to accept 47 submissions, including 17
papers in English and 30 in Chinese. We asked the authors of all the accepted papers to
submit a revised version based on the review reports.
This program would have not been possible without the efforts of the PC, the
external reviewers, and the authors. We would like to express our gratitude to all the
authors who submitted their papers. We would like to convey our deepest and sincerest
appreciation for all the hard work and dedication of our PC members and external
reviewers. We also gratefully acknowledge the kind support from our general chair,
Prof. Yong Dou, organization chair, Prof. Kuanjiu Zhou, and our Steering Committee.
Our thanks also go to the China Computer Federation (CCF), Technical Committee on
Computer Architecture of CCF, Dalian University of Technology, the City of Yinkou,
Xilinx, Baidu, and all the other institutes that kindly helped us. Finally, we greatly
appreciate the steady support provided by Springer.
General Chair
Yong Dou National University of Defense Technology, China
Organization Chair
Kuanjiu Zhou Dalian University of Technology, China
Program Chair
Chao Li Shanghai Jiao Tong University, China
Steering Committee
Zhenzhou Ji Harbin Institute of Technology, China
Chenggang Wu Institute of Computing Technology, CAS, China
Dongsheng Wang Tsinghua University, China
Junjie Wu National University of Defense Technology, China
Xingwei Wang Northeastern University, China
Gongxuan Zhang Nanjing University of Science and Technology, China
Program Committee
Quan Chen Shanghai Jiao Tong University, China
Zidong Du Institute of Computing Technology, CAS, China
Binzhang Fu Huawei
Yu Hua Huazhong University of Science and Technology, China
Weixing Ji Beijing Institute of Technology, China
Jingwen Leng Shanghai Jiao Tong University, China
Dongsheng Li National University of Defense Technology, China
Duo Liu Chongqing University, China
Yuhang Liu Institute of Computing Technology, CAS, China
Youyou Lu Tsinghua University, China
Guojie Luo Beijing University, China
Bo Mao Xiamen University, China
Songwen Pei University of Shanghai for Science and Technology, China
Minghua Shen Sun Yat-sen University, China
Wei Song Institute of Information Engineering, CAS, China
Guangyu Sun Beijing University, China
Jing Wang Capital Normal University, China
Lei Wang National University of Defense Technology, China
VIII Organization
Additional Reviewers
Qiang Cao Huazhong University of Technology, China
Li Jiang Shanghai Jiao Tong University, China
Naifeng Jing Shanghai Jiao Tong University, China
Cheng Li University of Science and Technology of China
Tao Li Nankai University, China
Yao Shen Shanghai Jiao Tong University, China
Shuang Song University of Texas at Austin, USA
Rui Wang Beihang University, China
Chentao Wu Shanghai Jiao Tong University, China
Qiaosha Zhou Zhejiang Sci-Tech University, China
Contents
Accelerators
EffectFace: A Fast and Efficient Deep Neural Network Model for Face
Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Weicheng Li, Dan Jia, Jia Zhai, Jihong Cai, Han Zhang, Lianyi Zhang,
Hailong Yang, Depei Qian, and Rui Wang
X Contents
Distributed Data Load Balancing for Scalable Key-Value Cache Systems. . . . 181
Shanshan Chen, Xudong Zhou, Guiping Zhou, and Richard O. Sinnott
1 Introduction
flows takes a lot of time to simulate and compile before actually running hard-
ware accelerators. With the development of FPGA high-level synthesis tool
(HLS), high-level programming language (C/C++) is used to replace low-level
HDL, which improves the speed of FPGA implementation and verification flows.
Greatly reducing the development cycle, to design FPGA has brought great con-
venience. In recent yeas, the use of HLS to design CNN architecture has contin-
ued to emerge. The work of [9] using the Vivado-HLS tool on a Xilinx VC707
FPGA board. However, only convolution layers are implemented on AlexNet [1].
In [10], author present a systematic methodology for maximizing the through-
put of an FPGA-based accelerator. In this work, an entire CNN model is pro-
posed consisting of all CNN layers: convolution, normalization, pooling and clas-
sification layers. The scalable of accelerator architecture only use like AlexNet
and VGG [4]. The feedforward neural networks with shortcut connections like
ResNet [5] dose not work. The main contribution of this work are:
2 Background
where Di (fi , y, x) and Do (fo , y, x) denote the neurons at position (x, y) in the
input feature map fi and output feature map fo , respectively. Wl (fo , fi , y, x) rep-
resents the corresponding weights in the l-th layer that gets convolved with fi .
The size of the convolution filters is K × K, while the total number of input fea-
ture maps is Cl . In addition to this, AlexNet considered the use of the ReLU non-
linearity instead of the saturating nonlinearites, such as sigmoids; Using dropout
in training and Local Response Normalization (LRN) to reduce the problem of
overfitting.
A Scalable FPGA Accelerator for Convolutional Neural Networks 5
VGG. VGG achieves its depth by simply stacking more layers while following
the standard practices introduced with AlexNet. The size of the convolution
kernel is more regular. AlexNet use 11 × 11, 5 × 5 and 3 × 3 filters, but VGG only
use 3 × 3 filters in the entire network. Notably, while using smaller filters, VGG
required far more filters per layer. The amount of calculations and parameters
of VGG is much larger than AlexNet.
ResNet. Deeper neural networks are more difficult to train, so in [5], author
proposed residual learning framework reducing the vanishing gradient problem
to ease the training of networks. This residual learning is mainly use of shortcut
connections, illustrated in Fig. 2, that connect components of different layers
with an identity mapping [6]. In particular, ResNet is built such that each layer
learns an incremental transformation, F (x), on top of the input, x, according to
vendors such as Xilinx and Intel have released OpenCL SDK for programming
FPGAs. The Intel OpenCL environment which can be a mixture of C, C++,
and OpenCL, provides a complete CPU/GPU-like development experience and
run-time experience on a CPU/FPGA platform, including a complete software
workflow spanning multiple target devices and x86 emulation with cycle-accurate
FPGA hardware models and cycle-accurate FPGA hardware.
(1) The cascaded and overlaped kernels form a deep pipeline architecture.
(2) Using a single hardware kernel to implement both the convolution and fully
connected layers.
(3) Scalable hardware structure which implementation many classic CNNs oper-
ations, such as LRN kernel to AlexNet, BatchNorm kernel and Eltwise kernel
to ResNet.
In this way, data vectorization and parallel CU structure are both explored
in the design. Vectorized input features Di and weights Wl are streamed by
multiple Channels. A design parameter VEC SIZE determines the degree of
data vectorization and controls the input throughput. Another design variable
parameter to accelerator the convolution operation CU NUM, represents the
parallel factor of weight and reuse factor of data. Due to efficient pipelined by
the OpenCL compiler, We propose an efficient convolution pipeline structure
consisted of a multiplier-adder tree with a delayed buffer as in Fig. 4.
Data Mover Kernel. Two multi-model single work-item kernels are designed
to fetch/store data from/to the global memory for the computation pipelines.
MemRD kernel detailed schemes in [11] can fetch data from global memory
to convolution mode or FC mode. We propose design parameter FT NUM to
determine the size of local memory, which further influences the reuse of input
data. MemWR kernel is mainly used to receive the output of convolution ker-
nel through the channel and arrange it into the storage structure required for
the next convolution or pooling operation. For the convolution mode, the data
received from the channel is arranged to have a depth of CU NUM, and MemWR
kernel need to divide the depth into VEC SIZE copies and return it to global
memory. The pooling mode simply transfer the data received from the channel
and directly put back to global memory. In the pooling mode, the MemWR ker-
nel also needs to pipe the synchronization signal to the pooling kernel at the right
time for them can overlap work. Note all memory operations should be commit-
ted before sending token to the pooling kernel. Detailed MemWR schemes are
illustrated in Fig. 5.
synchronization signal comes, the pooling kernel read data from global memory
to shift register. In the process of data transfer, if the time point of pooling
is reached, data will be extracted from the shift register to the pooling logic.
Similarly, we designed a parameter PT NUM for adjusting the size of the local
memory in pooling kernel to exploiting input data reuse. In the pooling strategy,
the first line is processed first, then the second line is compared with the first line,
and so on. The final result of the pooling calculation is stored in the ping-pong
buffer. During the pooling calculation, the result of the last calculation is also
divided into VEC SIZE and returned to global memory for the next convolution
calculation. This process is similar to MemWR.
Other Kernel. Besides the most compute-intensive convolution and fully con-
nected kernel, we also designed some common opencl kernels, such as LRN,
BatchNorm, Eltwise for the scalability and integrity of the CNN accelerator’s
overall design. In this architecture, you can choose the basic units used in the
network to piece together to implement different network structures. For exam-
ple, implementation AlexNet just choose convolution kernel, pooling kernel and
LRN kernel. Therefore, this scalable architecture can process the complete CNN
forword computation flow with little involvement of host CPU.
Other Layers Time. Due to the idea of pipeline and overlap in the overall
hardware design, the execution time of other kernels can be basically ignored
relative to convolution and fully connected operations.
reuse, the MemRD kernel design a FT NUM parameter and maxpool kernel
design a PT NUM parameter. These kernel fetches a window of data that cov-
ers the area of FT NUM or PT NUM of filters each time, and caches the data
in the on-chip buffers or shift register.
5 Experimental Results
In this section, we present the experimental results to validate the scalable of
this CNN accelerator by implementation three large-scale CNN models: AlexNet,
VGG-16 and ResNet-50 on DE5-Net platform.
We use DE5-Net FPGA development board from Altera and compare with
DE5a-Net listed its specification in Table 2. The OpenCL kernel codes are com-
piled using Altera OpenCL SDK 16.1, and the Quartus 16.1 is used as the FPGA
implementation tool. The host machine is equipped with an Intel i7-5930K CPU
and 64 GB memories. The data of images are first loaded from hard disks to
the host programs, and then sended to the FPGA accelerators to perform CNN
forword computations.
In this subsection, we first list the best parameter configuration on different net-
works. Then, we show the benchmark of our CNN accelerator. Finally, we dis-
cuss the scalability of this hardware architecture. As discussed in 4, four design
parameters VEC SIZE, CU NUM, FT NUM, PT NUM are used to control the
hardware cost and throughput of the FPGA accelerator. Therefore, design space
exploration can be quantitatively performed by implementing the accelerator
with different parameter configuration. The final design variables for three net-
works optimized on the DE5-Net board are shown in Table 3.
In Table 4, we summarize the resource utilization, execution time and per-
formance of different networks on the best parameters. We can see that different
networks have different parameters and achieve different performance on same
FPGA board. To prove how fast this accelerator can accelerate CNN computa-
tions, we also compare with CPU by using the Caffe deep learning framework.
The execution time for AlexNet, VGG-16 and ResNet-50 is 189 ms, 1547 ms
and 1238 ms, respectively. We can see that using FPGA-based accelerator can
achieve more than 10 times faster on average in implementation CNN-based
A Scalable FPGA Accelerator for Convolutional Neural Networks 13
6 Conclusion
In this work, we implemented a scalable FPGA accelerator for convolutional neu-
ral networks using OpenCL framework. An efficient and scalable hardware archi-
tecture with deep pipelined kernels was presented. We proposed and explored
four design parameters for hardware costs and bandwidth limited, and imple-
mented three large-scale CNNs, AlexNet, VGG-16 and ResNet-50 on DE5-Net
FPGA board.
References
1. Krizhevsky, A., Sutskever, I., Hinton, G.E., et al.: Imagenet classification with
deep convolutional neural networks. In: Advances in Neural Information Processing
Systems, pp. 1097–1105 (2012)
2. Ren, S., et al.: Faster R-CNN: towards real-time object detection with region pro-
posal networks. In: Advances in Neural Information Processing Systems, pp. 91–99
(2015)
14 K. Xu et al.
Zikai Nie, Zhisheng Li, Lei Wang(B) , Shasha Guo, and Qiang Dou
1 Introduction
Deep Neural Networks (DN N s) are Machine Learning (M L) methods that can
learn a generic but effective representation of an input space from large amount
of data. They extract high level features using previous learned models from raw
data to infer the final data distribution.
Over the past decade, DNNs, especially deep convolutional neural networks,
have gained a huge development due to the outstanding experimental achieve-
ments of multiple related fields including computer vision [14], speech recognition
[9], and natural language processing [8]. In many specific situations, DNNs used
in some domains are now able to beyond human in both accuracy and speed.
The success of DNNs can be attributed to three main reasons: the availabil-
ity of large-scale data sets, the developments of deep structures and training
c Springer Nature Singapore Pte Ltd. 2018
C. Li and J. Wu (Eds.): ACA 2018, CCIS 908, pp. 15–29, 2018.
https://doi.org/10.1007/978-981-13-2423-9_2
16 Z. Nie et al.
Fig. 1. Computation of a CONV/FC layer. Hi and Wi are the height and width of
ifmaps. Other notations are explained behind Eq. 1
3 Motivation
then there will be some problems. First, on-chip storage is not enough anymore,
and we have to leverage data reuse better to reduce the number of off-chip mem-
ory access and save energy. Second, with the deeper going of network structures,
there are more convolutional layers with small fmaps and filters. Third, the num-
ber of psums produced by parallel contexts is too large to make psums stay in
on-chip buffers.
In most widely used CNNs, such as LeNet [15], AlexNet [14] and VGG [20],
convolutional layers account for over 90% of the overall operations and produce
a large amount of data movement [5]. Thus, convolutional layers are important
for CNNs to gain high performance in throughput and energy efficiency.
There are two issues limiting throughput and energy efficiency of convolution.
First, a MAC operation that creates read requests of ifmaps and weights stored
in off-chip DRAM results in requirements of high bandwidth and energy con-
sumption. Second, a significant amount of partial sums (psums) are produced
by limited parallel contexts simultaneously, which introduce additional read or
write pressure and energy of access if not accumulated within an acceptable
time.
To deal with the first issue, we should leverage different types of data reuse:
– Sliding reuse. The value of S (stride) is always less than that of Fh and
Fw in convolutional layers to slow down the evolution roughening and to gain
more information from neighbors. This characteristic makes small amount of
ifmaps’ pixels become shared across many MAC kernels. Each pixel in ifmaps
can be used Fh × Fw (with padding) times in a 2D fmap with 2 directions.
– Ifmap reuse
– Intra-image filter reuse. According to Eq. 1, each 2D filter can be identified
by a couple of an input channel and an output channel. Therefore, convolu-
tional operations of an input channel use only one 2D filter to generate psums
of the same output channel. This kind of reuse doesn’t exist in FC layers.
– Inter-image filter reuse. Each filter can be further reused across the batch
of N ifmaps.
Fig. 2. The number of DRAM accesses of various buffer size. The 5 bars in each cluster
represent the 5 CONV layers of AlexNet. With various size of the on-chip buffers, we
crawl the output requests of the buffer and record the number of off-chip accesses.
“O Pericles,” said the dying exile, “those who need a lamp should
take care to supply it with oil!”
But how many lights of our latter-day lives have [181]thus been
extinguished before their time! Not one of the plethoric British
aristocrats who spiced their leisure with the sweets of poetry ever
dreamed of relieving the cruel distress of Robert Burns, or of cutting
the knot of the financial embroglio that strangled out the life of Sir
Walter Scott.
[Contents]
E.—REFORM.
[Contents]
A.—LESSONS OF INSTINCT.
[Contents]
B.—REWARDS OF CONFORMITY.
In the simple lives of the lower animals every day may bring the
sufficient reward of its toil; but the problem of progress, even from
the first dawn of civilization, involves tasks too apt to extend beyond
the span of individual existence. The forest-clearing husbandman,
the state-founding patriot, the scientific inquirer, all risk to receive the
summons of night before the completion of their labor. Before
reaching the goal of their hopes their earthly pilgrimage may end at
the brink of the unknown river, and education alone can bridge that
gulf, and make every day the way-station, of an unbroken road.
Children or children’s children will take up the staff from the last
resting-place of their pilgrim father; and, moreover, all progress is
cumulative. Every laborer works with the experience of his
forefathers, as well as his own; [186]every son stands on the
shoulders of his father. Even the failure of individual efforts
contributes a helpful lesson to the success of the next attempt:
and the vast fabric of our republican federation was founded by the
poor colonists who sought independence in the freedom of the
wilderness, and combined against the power of a selfish despot.
Education sows a seed which may sprout even during the life-time of
the sower, and bless individual life with the sweets of a guaranteed
triumph over the power of death. Resurgam, “I shall live after death,”
expresses the significance of that triumph, and of the “esoteric
doctrine of Pythagoras.”
[Contents]
C.—PERVERSION.
[Contents]
D.—PENALTIES OF NEGLECT.
Such was the morality which arrogated the right of suppressing that
system of physical and intellectual education which had filled the
homes of the Mediterranean nations with all the blessings of health,
science, and beauty. Theological training had failed to kindle the
dawn of a supernatural millennium, but had thoroughly succeeded in
extinguishing the light of human reason. Not absolute ignorance
only, but baneful superstition—worse than ignorance by just as much
as poison is worse than hunger—was for centuries the inevitable
result of all so-called school-training; and the traditions of that age of
priest-rule have made religion almost a synonyme of cant. It also
gave book-learning its supposed tendency to mental aberration. Can
we wonder at that result of an age when the literary products of
Christian Europe were confined almost exclusively to ghost-stories
and manuals of ceremony? Can we wonder that delusions of the
most preposterous kind assumed the virulence of epidemic
diseases? Maniacs of self-mutilation, of epileptic contortions, of
were-wolf panics, traversed Europe from end to end. Men gloried in
ignorance, and boasted their neglect of worldly science till the
consequences [190]of that neglect avenged its folly in actual
madness.
The saddest of all the sad “it might have beens” is, perhaps, a
reverie on the probable results of earlier emancipation—of the
employment of thirteen worse than wasted centuries in scientific
inquiries, agricultural improvements, social and sanitary reforms. We
might have failed to enter the portals of the New Jerusalem, but we
would probably have regained our earthly paradise.
[Contents]
E.—REFORM.
The days of the Holy Inquisition are past; but the restless
propaganda of Jesuitry still shames the inactivity of Rationalism. Our
friends sit listless, relying on the theoretical advantages of their
cause, while the busy intrigues of our enemies secure them all
practical advantages.
“If life shall have been duly rationalized by science,” says Herbert
Spencer, “parents will learn to consider a sound physical constitution
as an entailed estate, which should be transmitted unimpaired, if not
improved;” and with a similar recognition of social obligations
Freethinkers should endeavor to transmit to their children a bequest
of unimpaired common sense. Loyalty to their Protestant ancestors,
loyalty to posterity, and to the majesty of truth herself, should prompt
us to stand [193]bravely by our colors and train our children to
continue the struggle for light and independence.
[Contents]
CHAPTER XVI.
FOREST CULTURE.
[Contents]
A.—LESSONS OF INSTINCT.
[Contents]
B.—REWARDS OF CONFORMITY.
The love of forest-trees is a characteristic of the nature-abiding
nations of the North, and has rewarded itself by an almost complete
reversion of the original contrast between the garden lands of the
South and the inhospitable wilderness of the higher latitudes. Forest
destruction has turned Southern Europe into a desert, while the
preservation of forests has made the homes of the hyperborean
hunters an Eden of beauty and fertility. “One-third to the hunter, two-
thirds to the husbandman,” was the rule of Margrave Philip in his
distribution of forest and fields, and expresses the exact proportion
which modern science indicates as most favorable to the perennial
fertility of our farm-lands. In a single century the forest-destroying
Spaniards turned many of their American colonies from gardens into
sand-wastes, while, after fourteen hundred years of continuous
cultivation, the fields of the Danubian Valley are still as fertile as in
the days of Trajan and Tacitus. Along the river-banks and half-way
up the foot-hills the arable land has been cleared, but higher [197]up
the forest has been spared. All the highlands from Ratisbon to Buda-
Pesth still form a continuous mountain park of stately oaks and
pines, and, as a consequence, springs never fail; crops are safe
against winter floods and summer drouths; song-birds still return to
their birthland, and reward their protectors by the destruction of
noxious insects; meadows, grain-fields, and orchards produce their
abundant harvest year after year; famine is unknown, and
contagious diseases rarely assume an epidemic form. In Switzerland
and Prussia the preservation of the now remaining woodlands is
guaranteed by strict protective laws; Scandinavia requires her forest-
owners to replant a certain portion of every larger clearing; in Great
Britain the parks of the ancient mansions are protected like sacred
monuments of the past, and landowners vie in lining their field-trails
with rows of shade-trees. The fertility of those lands is a constant
surprise to the American traveler disposed to associate the idea of
eastern landscapes with the picture of worn-out fields. Surrounded
by Russian steppes and trans-Alpine deserts, the homes of the
Germanic nations still form a Goshen of verdure and abundance.
Forest protectors have not lost their earthly paradise.
[Contents]
C.—PERVERSION.