[Bioclusters] How can I make blast job running short time on Gridengine

bioinfo Gu bioclusters@bioinformatics.org
Thu, 21 Nov 2002 11:59:48 -0800 (PST)


--0-1162847587-1037908788=:14428
Content-Type: text/plain; charset=us-ascii


Hi Mike,
  Hi Grace,    A popular method for running jobs on two different machines at once is to divide the input into parts, send each part to a different machine, run to program on each machine using the segment of the input on that machine, then combining the results. This is what's usually called an embarrassingly parallel method, where each job has 5 parts:1) pre-processing (preparing data) on the submission host2) data transfer to the nodes3) processing on the execution nodes4) data transfer back to the submission node, queuing of the results5) post-processing (combining the results) on the submission host But this should be done through a lot of work, and make some scripts, right? So, for the case where you are using BLAST as the application, the database (or query) can be split on sequence boundaries, sent to each of the nodes for BLASTing, result files sent back to the submission host and combined to get the final result. This yields about 1/N performance depending on the ef
 ficiency of your configuration, where N is the number of nodes in your cluster. This would be one way of getting roughly 2x the performance our of your SGE cluster than what you would get out of a single machine.     Another nice feature of some Distributed Resource Management (DRM) tools like (LSF, SGE ...) on clusters is that they do some level of load balancing. So, if you needed to run a job and needed it to have a fair chance of getting run with whatever everyone else is doing, the DRM would figure out which machine will give you the best service for your job. One nice feature of some schedulers in DRM packages (and they are not all equal!!) is that each user, group, job... can have a priority placed on it that will actually preempt other jobs, shuffle queuing... to get the right resources into the hands of the people who need them most..  I have no cluser in hand, I just install SGE on two separate unix machines, so far, I can not think SGE can shorten certain job's ru
 nning time, it just distribute the job to idle host, of course, if one host is too busy, it do not need wait, which will avaoid the job packed on some hosts. But suppose all the host are idle, the submitted job still running on one host. So I want to know if there is big efficience difference between a collection of computer and real cluster. Look forward your suggestion. Grace Combining parallel (embarrassingly parallel) job execution with scheduling/load-balancing features of DRM tools is really the key to achieving the efficiency in a cluster that makes if a valuable resource for doing things like BLAST. ________________________________________ 
Mike McCardle 
Systems Software Engineer 
RLX Technologies, Inc. 
mike.mccardle@rlxtechnologies.com 
http://www.rlxtechnologies.com 

From: bioinfo Gu [mailto:bioinfowistar@yahoo.com]
Sent: Thursday, November 21, 2002 10:08 AM
To: bioclusters@bioinformatics.org
Subject: [Bioclusters] How can I make blast job running short time on Gridengine



Hi all,

I have two machines to install SGE. athena is master and excution host, apollo is only excution host. When I submit a job from master host with qsub, the job will be distributed to one of queue(one of host), and this job will be executed on this machinefor the whole process. I can not see Gridengine can save execution time when launch blast job on it.  how can I save blast running time on Gridengine, do I have to use Parallel Environment?

Also, how can I setup environmental variable for specific execution host in batch job script? For example: 

on one excution host: I need BLASTDB point to path1, on the second execution host, I want to set BLASTDB to path2, how can I do that?

Thank you very much in advance.

Grace



---------------------------------
Do you Yahoo!?
Yahoo! Mail Plus - Powerful. Affordable. Sign up now


---------------------------------
Do you Yahoo!?
Yahoo! Mail Plus - Powerful. Affordable. Sign up now
--0-1162847587-1037908788=:14428
Content-Type: text/html; charset=us-ascii

<P>Hi Mike,
<P><B><I>&nbsp;<A href="mailto:mike.mccardle@rlx.com">&nbsp;</A></I></B>
<META content="MSHTML 5.50.4522.1800" name=GENERATOR>
<BLOCKQUOTE style="PADDING-LEFT: 5px; MARGIN-LEFT: 5px; BORDER-LEFT: #1010ff 2px solid">
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>Hi Grace,</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>&nbsp;&nbsp;&nbsp; A popular method for running jobs on two different&nbsp;machines at once is to divide the input into parts, send each part to a different machine, run to program on each machine using the segment of the input on that&nbsp;machine, then combining the results. This is what's usually called an embarrassingly parallel method, where each job has 5 parts:</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>1) pre-processing (preparing data) on the submission host</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>2) data transfer to the nodes</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>3) processing on the execution nodes</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>4) data transfer back to the submission node, queuing of the results</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>5) post-processing (combining the results) on the submission host</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002></SPAN></FONT>&nbsp;</DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>But this should be done through a lot of work, and make some scripts, right?</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>&nbsp;</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>So, for the case where you are using BLAST as the application, the database (or query) can be split on sequence boundaries, sent to each of the nodes for BLASTing, result files sent back to the submission host and combined to get the final result. This yields about 1/N performance depending on the efficiency of&nbsp;your configuration, where N is the number of nodes in your cluster.&nbsp;This would be one way of getting roughly 2x the performance our of your SGE cluster than what you would get out of a single machine.</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002></SPAN></FONT>&nbsp;</DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>&nbsp;&nbsp;&nbsp;&nbsp;Another&nbsp;nice feature of some Distributed Resource&nbsp;Management (DRM) tools like (LSF, SGE ...) on clusters is that they do some level of load balancing. So, if you needed to run a job and&nbsp;needed it to have a fair chance of getting run&nbsp;with whatever everyone else is doing, the DRM would figure out which machine will give you the best service for your job. One nice feature of some schedulers in DRM packages (and they are not all equal!!) is that each user, group, job... can have a priority placed on it that will actually preempt other jobs, shuffle queuing... to get the right resources&nbsp;into the hands&nbsp;of the&nbsp;people who need them most..&nbsp;</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002></SPAN></FONT>&nbsp;</DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>I have no cluser in hand, I just install SGE on two separate unix machines, so far, I can not think SGE can shorten certain job's running time, it just distribute the job to idle host, of course, if one host is too busy, it do not need wait, which will avaoid the job packed on some hosts.&nbsp;But suppose all the host are idle, the submitted job still running on one host.&nbsp;So I want to know if there is big efficience difference between a collection of computer and real cluster.</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002></SPAN></FONT>&nbsp;</DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>Look forward your suggestion.</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002></SPAN></FONT>&nbsp;</DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>Grace</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>&nbsp;</SPAN></FONT></DIV>
<DIV><FONT face=Arial color=#0000ff size=2><SPAN class=754271417-21112002>Combining parallel (embarrassingly parallel)&nbsp;job execution with&nbsp;scheduling/load-balancing&nbsp;features of DRM tools&nbsp;is really the key to achieving the efficiency&nbsp;in a cluster that makes if a valuable resource for doing things like BLAST. </SPAN></FONT></DIV>
<DIV><SPAN class=754271417-21112002><FONT face=Arial><FONT color=#0000ff size=2>________________________________________ </FONT></FONT></DIV>
<DIV>
<P><FONT face=Arial><FONT color=#0000ff><FONT size=2>Mike McCardle <BR>Systems&nbsp;<SPAN class=754271417-21112002>Software </SPAN>Engineer <BR>RLX Technologies, Inc. <BR><A href="mailto:mike.mccardle@rlxtechnologies.com">mike.mccardle@rlxtechnologies.com</A> <BR></FONT><A target=_blank href="http://www.rlxtechnologies.com/"><FONT size=2>http://www.rlxtechnologies.com</FONT></A><FONT size=2> </FONT></FONT></FONT></P>
<P></SPAN><FONT face=Tahoma size=2><B>From:</B> bioinfo Gu [mailto:bioinfowistar@yahoo.com]<BR><B>Sent:</B> Thursday, November 21, 2002 10:08 AM<BR><B>To:</B> bioclusters@bioinformatics.org<BR><B>Subject:</B> [Bioclusters] How can I make blast job running short time on Gridengine<BR><BR></FONT></P></DIV>
<BLOCKQUOTE>
<P>Hi all,</P>
<P>I have two machines to install SGE. athena is master and excution host, apollo is only excution host. When I submit a job from master host with qsub, the job will be distributed to one of queue(one of host), and this job will be executed on this machinefor the whole process. I can not see Gridengine can save execution time when launch blast job on it.&nbsp; how can I save blast running time on Gridengine, do I have to use Parallel Environment?</P>
<P>Also, how can I setup environmental variable for specific execution host in batch job script? For example: </P>
<P>on one excution host: I need&nbsp;BLASTDB point to path1, on the second execution host, I want to set&nbsp;BLASTDB to path2, how can I do that?</P>
<P>Thank you very much in advance.</P>
<P>Grace</P>
<P><FONT face=Arial color=#0000ff size=2></FONT><FONT face=Arial color=#0000ff size=2></FONT><FONT face=Arial color=#0000ff size=2></FONT><FONT face=Arial color=#0000ff size=2></FONT><FONT face=Arial color=#0000ff size=2></FONT><BR>
<HR SIZE=1>
Do you Yahoo!?<BR><A href="http://rd.yahoo.com/mail/mailsig/*http://mailplus.yahoo.com">Yahoo! Mail Plus</A> - Powerful. Affordable. <A href="http://rd.yahoo.com/mail/mailsig/*http://mailplus.yahoo.com">Sign up now</A></BLOCKQUOTE></BLOCKQUOTE><p><br><hr size=1>Do you Yahoo!?<br>
<a href="http://rd.yahoo.com/mail/mailsig/*http://mailplus.yahoo.com">Yahoo! Mail Plus</a> - Powerful. Affordable. <a href="http://rd.yahoo.com/mail/mailsig/*http://mailplus.yahoo.com">Sign up now</a>
--0-1162847587-1037908788=:14428--