Terri,

Per our discussion today, the logic needs to be modified slightly to
accommodate bootstopping, i.e., when -N autoyy is specified, where
autoyy is not a number.  The changes are flagged in the appended
message by **s.

Regards,  Wayne

>Date: Fri, 11 Dec 2009 11:16:14 -0800
>To: terri@sdsc.edu
>From: Wayne Pfeiffer <pfeiffer@sdsc.edu>
>Subject: Fwd: Hybrid parallel version of RAxML is available for production
>  use
>Cc: mmiller@sdsc.edu, pfeiffer@sdsc.edu
>
>Hi Terri,
>
>Here is an update from my earlier e-mail message regarding the
>recommended number of nodes, MPI processes, and Pthreads when
>running RAxML 7.2.5 on Abe.

When -N nnn -x xxx -f a is specified **with nnn a number**, check the
number of characters (nchar)

>in the PHYLIP file, or ask the user to input that separately.

When -N nnn is specified **with nnn a number**, the hybrid code code is

>recommended for all but one Pthreads-only case per the following:
>
>   For nnn < 20, use one node with
>     . 8 Pthreads (no MPI) if -xxx -f a is specified & nchar >= 1,000 or
>     . 2 MPI processes & 4 Pthreads otherwise.
>   For nnn >= 20 & nnn < 50, use 2 nodes with
>     . 2 MPI processes & 8 Pthreads if -xxx -f a is specified &
>nchar >= 1,000 or
>     . 4 MPI processes & 4 Pthreads otherwise.
>   For nnn >= 50 & nnn < 100, use 5 nodes with
>     . 5 MPI processes & 8 Pthreads if -xxx -f a is specified &
>nchar >= 1,000 or
>     . 10 MPI processes & 4 Pthreads otherwise.
>   For nnn >= 100, use 10 nodes with
>     . 10 MPI processes & 8 Pthreads if -xxx -f a is specified &
>nchar >= 1,000 or
>     . 20 MPI processes and 4 Pthreads otherwise.
>
>For all other options to be run on Abe, use 8 Pthreads in one node.
>
>As noted before, two versions of the executable need to be built:
>one for Pthreads only and the other for hybrid MPI/Pthreads.
>
>* Let me know how much effort will be required to implement this and
>when you and Mark think you can do it.
>
>Thanks,  Wayne
>
>>Date: Mon, 30 Nov 2009 10:20:32 -0800
>>To: Mark Miller <mmiller@sdsc.edu>, terri@sdsc.edu
>>From: Wayne Pfeiffer <pfeiffer@sdsc.edu>
>>Subject: Hybrid parallel version of RAxML is available for production use
>>Cc: pfeiffer@sdsc.edu
>>
>>Hi Mark & Terri,
>>
>>Recently Alexis made the hybrid parallel capability available in
>>RAxML 7.2.4.  Subsequently he added optimizations for parsimony
>>functions and consensus tree building in RAxML 7.2.5.
>>
>>Thus I recommend that we make RAxML 7.2.5 available instead of
>>7.2.3. This could provide users with *very* significant speedups
>>over 7.0.4, which still can be supported for now.
>>
>>The choice of the number of nodes, MPI processes, and Pthreads is
>>more complicated with 7.2.5 as compared to 7.2.3.  Here is what I
>>suggest, all for runs on Abe.
>>
>>-N nnn (but no -x xxx -f a)
>>
>>   For nnn < 10, use 8 Pthreads in one node.
>>   For nnn >= 10 & nnn < 20, use 5 MPI processes and 8 Pthreads in five nodes.
>>   For nnn >= 20, use 10 MPI processes and 4 Pthreads in five nodes.
>>
>>-x xxx -f a -N nnn
>>
>>   For nnn < 50, use 8 Pthreads in one node.
>>   For nnn >= 50 & nnn < 100, use 5 MPI processes and 8 Pthreads in
>>five nodes.
>>   For nnn >= 100, use 10 MPI processes and 4 Pthreads in five nodes.
>>
>>For all other options to be run on Abe, use 8 Pthreads in one node.
>>
>>* Let me know what you think of this.
>>
>>Regards,  Wayne
>>
>>P.S.  Two versions of the executable need to be built: one for
>>Pthreads only and the other for hybrid MPI/Pthreads.