Newer
Older
\documentclass[a4paper]{jpconf}
\usepackage[english]{babel}
% \usepackage{cite}
\usepackage{biblatex}
%\bibliographystyle{abnt-num}
%%%%%%%%%% Start TeXmacs macros
\newcommand{\tmtextit}[1]{{\itshape{#1}}}
\newenvironment{itemizedot}{\begin{itemize} \renewcommand{\labelitemi}{$\bullet$}\renewcommand{\labelitemii}{$\bullet$}\renewcommand{\labelitemiii}{$\bullet$}\renewcommand{\labelitemiv}{$\bullet$}}{\end{itemize}}
%%%%%%%%%% End TeXmacs macros
\begin{document}
\title{Evaluating Migration of INFN--T1 from CREAM-CE/LSF to
HTCondor-CE/HTCondor}
\author{Stefano Dal Pra$^1$}
\address{$^1$ INFN-CNAF, viale Berti-Pichat 6/2, 40127 Bologna, Italy}
\ead{stefano.dalpra@cnaf.infn.it}
\begin{abstract}
The Tier 1 datacentre provides computing resources for a variety of HEP and
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
Astrophysics experiments, organized in Virtual Organization submitting their
jobs to our computing facilities through Computing Elements, acting as Grid
interfaces to the Local Resource Manager. We planned to phase--out our
current LRMS (IBM/Platform LSF 9.1.3) and CEs (CREAM) to adopt HTCondor as a
replacement of LSF and HTCondor--CE instead of CREAM. A small cluster has
been set up to practice the management and evaluate a migration plan to a
new LRMS and CE set. This document reports about our early experience on
this.
\end{abstract}
\section{Introduction}
The INFN-T1 currently provides a computing power of about 400KHS06, 35000
slots on one thousand physical Worker Nodes. These resources are accessed
through Grid by 24 Grid VOs and locally by 25 user groups.
The IBM/Platform LSF 9.1.3 batch system arbitrate access to all the competing
users groups, both Grid and local, according to a \tmtextit{fairshare} policy,
designed to prevent underutilization of the available resources or starvation
of lower priority groups, while ensuring a medium--term share proportional to
configured quotas.
The CREAM--CEs act as frontend for Grid users to the underlying LSF batch
system, submitting their jobs on behalf of them. This setup has proven to be
an effective solution for several years. However, the compatibility between
CREAM and HTCondor seems to be less tight than with LSF. Moreover, active
development of CREAM has recently ceased and thus we cannot expect new
versions to be released, nor better HTCondor support to be implemented by an
officiale development team. We decided to migrate our batch system solution
from LSF to HTCondor, thus we need to also change our CEs. We have selected
HTCondor-CE as a natural choice, because it is maintained by the same
development team of HTCondor. Following we provide a report about our
experience with HTCondor and HTCondor--CE.
\section{The HTCondor cluster}
To get acquainted with the new batch system and CEs, to evaluate how these can
work together, how other components, such as monitoring, provisioning and
accounting systems can be integrated with HTCondor and HTCondor--CE and
finally to devise a reasonable migration plan, a simple small HTCondor 8.6.13
cluster has been set up during spring 2018. A HTCondor--CE was soon added, in
late April. HTCondor is a very mature opensource product, deployed at several
major Tier 1 for years, thus we already know that it will certainly fit our
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
use cases. The HTCondor--CE, on the other hand, is a more recent product, and
a number of issues might be too problematic for us to deal with. Our focus is
about ensuring that this CE implementation can be a viable solution for us.
\subsection{The testbed}
The test cluster consists of:
\begin{itemizedot}
\item a HTCondor-CE on top of
\item a HTCondor \ Central Manager and Collector
\item 3 Worker Nodes (Compute Nodes, in HTCondor terms), 16 slot each.
\end{itemizedot}
\subsection{HTCondor--CE Installation and setup}
The first CE installation was a bit tricky. The RPMs were available from OSG
repositories only, meaning that a number of default settings and dependencies
were unmet for EGI standards. Short after however, HTCondor--CE RPMs were made
available on \ the same official repository of HTCondor.
\subsubsection{Setup}
To setup the configuration for the HTCondor and HTCondor--CE puppet modules
are available. Unfortunately the puppet system at our site is not compatible
with these modules as they depend on \tmtextit{hiera}, which is not supported
at our site. These were later adapted to make them compatible with our
configuration management system. In the meanwhile, the setup was finalized
looking at the official documentation.
\subsubsection{Configuration.}
The first configuration was completed manually. The main documentation
source for the HTCondor--CE is that of the OSG website~\cite{OSGDOC},
which refers to a tool \tmtextit{osg-configure} not present on the
general HTCondor--CE release. Because of this the setup was completed
by trial and error. Once a working setup was obtained, a set of
integration notes were added to a public wiki~\cite{INFNWIKI}. This
should help other non OSG users to get some supplementary hint to
complete their installation.
\subsubsection{Accounting}
As of 2018, the official EGI accounting tool, APEL~\cite{APEL}, has no support for
HTCondor--CE. On the other hand, INFN--T1 has a custom accounting tool in
place for several years now~\cite{DGAS}. Thus, it's all about finding a suitable way to
retrieve from HTCondor the same information that we retrieve from CREAM--CE
and LSF.
A working way to do so has been by using python and the \tmtextit{python
bindings}, a set of api interfaces to the HTCondor daemons. These can be used
to query the SCHEDD at the CE and retrieve a specified set of data\quad about
recently finished jobs, which are subsequently inserted to our local
accounting database. A noticeable fact to note, is that the grid information
(User DN, VO, etc.) are directly available together with all the needed
accounting data. This simplifies the accounting problem, as it is no more
necessary to collect grid data separately from the BLAH component and then
look for matches with the corresponding grid counterpart.
This solution have been used during the 2018 to provide accounting for
HTCondor--CE testbed cluster.
\subsection{Running HTCondor--CE}
After some time to become confident with the main configuration tasks the
testbed begun working with jobs submitted by the 4 LHC experiments from
September 2018. The system proved to be stable and smooth, being able to work
unattended. This confirms that this system can be a reliable substitute for
CREAM--CE and LSF.
\subsection{Running HTCondor}
The HTCondor batch system is a mature product with a large user base. We have
put less effort at investigating it deeply. We already know that most or all
of needed features will work well. Rather, some effort have been put on
dealing with configuration management.
\subsubsection{Configuration management}
Eventhoutgh a standard base of puppet classes have been adapted to our
management system, an additional python tool have been written to improve
flexibility and readiness. The tool works by reading and enforcing on each
node of the cluster a set of configuration directives written on text files
accessible from a shared filesystem. The actual set and the read order depends
on the host role and name. Doing so, a large cluster can be quite easily
managed as a collection of set of host sets. The tool is quite simple and
limited but it can be improved as needed when more complex requirements should
arise.
\subsection{The migration plan}
After using the testbed cluster a possible plan for a smooth migration have
been devised:
\begin{itemizedot}
\item Install and setup a new HTCondor cluster, with a few more HTCondor--CE
and an initial small set of Worker Nodes
\item Enable the LHC VOs on the new cluster
\item Add more WN to the new cluster gradually
\item Enable other Grid VOs
\item Finally, enable submission from local submissions. These are made from
a heterogenous set of users, with a potentially rich set of individual needs
and can require a considerable administrative effort to meet all of them.
\end{itemizedot}
\subsection{Conclusion}
A testbed cluster based on HTCondor--CE on top of HTCondor batch system has
been deployed to evaluate these as a substitute for CREAM--CE and LSF. The
evaluation as mostly focused on the HTCondor--CE, as it is the most recent
product. Apart for a few minor issues, mainly related to gaps in the available
documentation, The CE proved to be a stable component. the possibility to
perform accounting has been verified.
Using the described testbed we have
\section*{References}
\begin{thebibliography}{9}
\bibitem{OSGDOC} \url{https://opensciencegrid.org/docs/compute-element/install-htcondor-ce/}
\bibitem{INFNWIKI} \url{http://wiki.infn.it/progetti/htcondor-tf/htcondor-ce_setup}
\bibitem{DGAS} S. Dal Pra, ``Accounting Data Recovery. A Case Report from
INFN-T1'' Nota interna, Commissione Calcolo e Reti dell'INFN, {\tt CCR-48/2014/P}
\bibitem{APEL} \url{https://wiki.egi.eu/wiki/APEL}
\end{thebibliography}
\end{document}