ptopgen.in

#!/usr/bin/perl -w

#
# EMULAB-COPYRIGHT
# Copyright (c) 2000-2003 University of Utah and the Flux Group.
# All rights reserved.
#

use English;
use Getopt::Std;

sub usage()
{
    print("Usage: ptopgen [-v] [-s switch] [-p pid [-e eid]] [-m factor]\n" .
	  "       -p include nodes the project has permission to use\n".
	  "       -e include given experiments resources\n" .
	  "          in the ptopfile (as if they were free)\n" .
	  "       -v Include stuff for topologies with virtual nodes\n".
	  "       -r Include stuff for topologies with widearea nodes\n".
	  "       -s Include stuff for topologies with simulated nodes\n".
	  "       -m Override multiplex_factor\n");
    exit(-1);
}
my $optlist = "s:e:m:vp:rS";
my $mfactor;
my $virtstuff = 0;
my $widearea  = 0;
my $simstuff  = 0;

#
# Turn off line buffering on output
#
$| = 1;

# Settings - Change these to your particular network.

#
# Also bogus. The interfaces table does not hold entries for the switch
# side of each link. It will have to at some point, when we have something
# other than 100Mbs ports (say, gbit ports).
#
# Speed in in Kbs!
#
my $default_switchport_speed    = 100000;

#
# Yet more bogosity - we hardcode this weight, which is given to wide-area
# (primarily plab) nodes to prefer spreading across sites
#
my $site_weight = 0.99;

######################################################################

my $TBROOT = "@prefix@";
use lib '@prefix@/lib';
require exitonwarn;
use libdb;

my $TRACK_INTERSWITCH_BANDWIDTH = "@TRACK_INTERSWITCH_BANDWIDTH@";

my %switches	  = ();
my %used_switches = ();
my %permissions   = ();
my %typemap       = ();
my %auxtypemap    = ();

my $DEADPID = NODEDEAD_PID();
my $DEADEID = NODEDEAD_EID();

my $pid;
my $exempt_eid;
my $switchtouse;

#
# Parse command arguments. Once we return from getopts, all that should be
# left are the required arguments.
#
%options = ();
if (! getopts($optlist, \%options)) {
    usage();
}
if (@ARGV) {
    usage();
}
if (defined($options{"s"})) {
    $switchtouse = $options{"s"};
}
if (defined($options{"m"})) {
    $mfactor = $options{"m"};
}
if (defined($options{"v"})) {
    $virtstuff = 1;
}
if (defined($options{"r"})) {
    $widearea = 1;
}
if (defined($options{"S"})) {
    $simstuff = 1;
}
if (defined($options{"p"})) {
    $pid = $options{"p"};
}
if (defined($options{"e"})) {
    $exempt_eid = $options{"e"};
    usage()
	if (!defined($pid));
}

# Read class/type maps
my $result =
    DBQueryFatal("select class,type,delay_capacity,".
		 "   virtnode_capacity,isvirtnode,simnode_capacity, " .
		 "   speed, RAM ".
		 "from node_types");
while (($class,$type,$delaycapacity,
	$virtcapacity,$isvirt,$simcapacity, $speed, $ram)
	    = $result->fetchrow_array) {
    $map = {};
    $map->{'CLASS'}    = $class;
    $map->{'DELAYCAP'} = $delaycapacity;
    $map->{'VIRTCAP'}  = $virtcapacity;
    $map->{'SIMCAP'}   = $simcapacity;
    $map->{'ISVIRT'}   = $isvirt;
    $map->{'SPEED'}    = $speed;
    $map->{'RAM'}      = $ram;
    $map->{'FEATURES'} = [];
    $map->{'AUXTYPES'} = {};
    $typemap{$type} = $map;

    # Default is permission to use all types,classes. 
    $permissions{$class} = 1;
    $permissions{$type}  = 1;
}

#
# Read the features table for each type.
# 
$result = DBQueryFatal("select type, feature, weight from node_type_features");
while (my ($type, $feature, $weight) = $result->fetchrow()) {
    push(@{$typemap{$type}->{'FEATURES'}}, "$feature:$weight");
}

#
# Read the auxtypes for each type.
# 
$result = DBQueryFatal("select auxtype,type from node_types_auxtypes");
while (my ($auxtype,$type) = $result->fetchrow()) {
    $typemap{$type}->{'AUXTYPES'}->{$auxtype} = 1;
    $auxtypemap{$auxtype} = $type;
}

#
# Read the features table for each individual node
#
$result = DBQueryFatal("select node_id, feature, weight from node_features");
while (my ($node_id, $feature, $weight) = $result->fetchrow()) {
    if (! defined($features{$node_id})) {
	$features{$node_id} = ["$feature:$weight"];
	next;
    } else {
	push @{$features{$node_id}}, "$feature:$weight";
    }
}
#
# Read in the node_auxtypes table for each node.
#
$result = DBQueryFatal("select node_id, type, count from node_auxtypes");
while (my ($node_id, $type, $count) = $result->fetchrow()) {
    if (! defined($auxtypes{$node_id})) {
	$auxtypes{$node_id} = ["$type:$count"];
	next;
    }
    push @{$auxtypes{$node_id}}, "$type:$count";
}
#
# Read the permission table if given a pid. If there is an entry in
# the table for a type/class, then permission is granted only if there
# is a record with the pid. If not, remove the permission granted above.
# 
if (defined($pid)) {
    $result =
	DBQueryFatal("select type from nodetypeXpid_permissions");
    
    while (my ($type) = $result->fetchrow_array) {
	$permissions{$type} = 0;
    }
    
    $result =
	DBQueryFatal("select type from nodetypeXpid_permissions ".
		     "where pid='$pid'");
    
    while (my ($type) = $result->fetchrow_array) {
	$permissions{$type} = 1;
    }
}

# Print switches
if (defined($switchtouse)) {
    print "node $switchtouse switch:1 *lan:*\n";
    $switches{$switchtouse} = 1;
}
else {
    $result =
	DBQueryFatal("select node_id from nodes where role = \"testswitch\"");

    while (($switch) = $result->fetchrow_array) {
	print "node $switch switch:1 *lan:*\n";
	$switches{$switch} = 1;
    }
}

# Find available nodes.
#
# This first query deals with just local nodes. Local nodes can host
# virtnodes, according to the delay_capacity in the types table. 
#

# the ordinary free condition for a local node.
my $free_condition = "(b.node_id is null and ".
                     " (np.eventstate='" . TBDB_NODESTATE_ISUP . "' or ".
                     "  np.eventstate='" . TBDB_NODESTATE_PXEWAIT . "')) ";

# if the user has specified an exempt pid/eid, 
# then view any node which is reserved to that experiment as available.
if (defined($exempt_eid)) {
    $free_condition = "($free_condition or ".
	"(b.pid='$pid' and b.eid='$exempt_eid'))"; 
}

$result =
    DBQueryFatal("select a.node_id,a.type,a.phys_nodeid,t.class,t.issubnode," .
		 "a.def_boot_osid from nodes as a ".
		 "left join reserved as b on a.node_id=b.node_id ".
		 "left join reserved as m on a.phys_nodeid=m.node_id ".
		 "left join nodes as np on a.phys_nodeid=np.node_id ".
		 "left join node_types as t on t.type=a.type ".
		 "where $free_condition and ".
		 "      (a.role='testnode' and t.isremotenode=0)");

#
# Scan the results, checking permissions and adding to the list
# You get to use a node type if no pid was specified (that is, you get
# to use all nodes), or if there is no entry in the perms table for
# the type/class of node.
#
while (($node,$type,$physnode,$class,$issubnode,$def_boot_osid) 
    = $result->fetchrow_array) {
    $nodes{$node} = $type
	if (!defined($pid) ||
	    ($permissions{$type} && $permissions{$class}));
    if ($issubnode) {
	$subnode_of{$node} = $physnode;
    }
    $node_def_boot_osid{$node} = $def_boot_osid;
}

foreach $node (keys(%nodes)) {
    my $type  = $nodes{$node};
    my $class = $typemap{$type}->{'CLASS'};
    my $delay_capacity = $typemap{$type}->{'DELAYCAP'};
    my $simnode_capacity = $typemap{$type}->{'SIMCAP'};
    my $cpu_speed = $typemap{$type}->{'SPEED'};
    my $ram = $typemap{$type}->{'RAM'};
    
    my @types = ("$type:1");
    my @features;
    my @flags;
    my $needvirtgoo = 0;

    # XXX temporary hack until node reboot avoidance 
    # is available. Nodes running the FBSD-NSE image
    # will have a feature def-osid-fbsd-nse 0.0
    # This is used by assign to prefer these pnodes
    # first before using others.
    if( $node_def_boot_osid{$node} eq 
	TBOSID(TB_OPSPID, "FBSD-NSE") ) { 
	push(@features, "FBSD-NSE:0.0");
    }

    # Might be equal, which assign would sum as two, not one!
    if ($type ne $class) {
	push(@types, "$class:1");
    }

    if ($delay_capacity > 0) {
	push @types, "delay:$delay_capacity";
    }

    #
    # Add any auxiliary types
    #
    foreach my $auxinfo (@{$auxtypes{$node}}) {
	my ($auxtype,$count) = split(":", $auxinfo);
	my $realtype;

	# Map an auxtype back to its real type, unless it is a real type.
	if (defined($auxtypemap{$auxtype})) {
	    $realtype = $auxtypemap{$auxtype};
	}
	else {
	    $realtype = $auxtype;
	}

	if ($typemap{$realtype}->{'ISVIRT'} && $count > 0) {
	    next
		if (! $virtstuff);
		
	    if (defined($mfactor) && $mfactor <= $count) {
		$auxinfo = "$auxtype:$mfactor";
	    }
	    else {
		$auxinfo = "$auxtype:$count";
	    }
	    $needvirtgoo = 1;
	}
	push(@types, $auxinfo);
    }

    #
    # This stuff is extra goo for local virtual nodes.
    # 
    if ($needvirtgoo) {
	push @types, "*lan:*";
	# Add trivial bw spec.
	push @flags, "trivial_bw:400000";
	# Add CPU and RAM information
	push @features, "?+cpu:$cpu_speed";
	push @features, "?+ram:$ram";
	# This number can be use for fine-tuning packing
	push @features, "?+virtpercent:100";
    }

    if ($simstuff && $simnode_capacity > 0) {
	#
	# Use user specified multiplex factor
	#
	if (defined($mfactor) && $mfactor <= $simnode_capacity) {
	    push @types, "sim:$mfactor";
	}
	else {
	    push @types, "sim:$simnode_capacity";
	}
	push @types, "*lan:*";
	# Add trivial bw spec.
	push @flags, "trivial_bw:100000";
    }

    # Add features
    push(@features, @{$typemap{$type}->{'FEATURES'}});
    if (defined($features{$node})) {
	push @features, @{$features{$node}};
    }

    #
    # Handle subnodes
    #
    if ($subnode_of{$node}) {
	# We don't want to include subnodes unless their parent node is going
	# to be in the ptop file too
	if (!$nodes{$subnode_of{$node}}) {
	    # In fact, nuke it from %nodes so that we don't include its links,
	    # either
	    delete $nodes{$node};
	    next;
	}
	# Push the subnode's information into its flags
	push @flags, "subnode_of:$subnode_of{$node}";
    }

    my $text = "node $node " . join(" ",@types) . " - " . join(" ",@features) .
	" - " . join(" ",@flags);


    print "$text\n";
}

#
# Widearea Nodes. Includes plab nodes. Note that widearea nodes are never
# allocated directly (they are in a holding experiment), but assign deals
# with it by allocating multiple vnodes on a pnode.
#
# The underlying physnode has to be "up", as determined by the
# autostatus stuff; this will prevent us from allocating a dead
# virtual node to an experiment.  This is especially hacky. We need
# another mechanism for this. We only add virtnodes when assign says
# we need them. This reduces the problem size for assign.
#
# The types we lay out are only those in the auxtypes table for the node,
# since that is where we define what vtypes are hosted on a particular
# physnode. 
#
if ($widearea) {
    $result =
	DBQueryFatal("select count(a.node_id),a.phys_nodeid,aa.type, ".
		     "  ns.status, wn.site ".
		     "  from nodes as a ".
 		     "left join reserved as b on a.node_id=b.node_id ".
		     "left join nodes as aa on aa.node_id=a.phys_nodeid ".
		     "left join reserved as m on a.phys_nodeid=m.node_id ".
		     "left join node_status as ns on ".
		     "     a.phys_nodeid=ns.node_id ".
		     "left join node_types as t on t.type=a.type ".
		     "left join widearea_nodeinfo as wn on ".
		     "     wn.node_id=a.phys_nodeid ".
		     "where  (a.role='virtnode' and t.isremotenode=1 and ".
		     "        (b.node_id is null and ".
		     "         !(m.pid='$DEADPID' and m.eid='$DEADEID'))) ".
		     "group by a.phys_nodeid");
    
    while (($count,$physnode,$ptype,$status,$site)
	    = $result->fetchrow_array) {
	my $class = $typemap{$ptype}->{'CLASS'};
	my @types;
	my @features;
	my @flags;

	# capacity is smaller of mfactor, number of unallocated virtual node
	# and the per-auxtype capacity below.
	my $virtnode_capacity = $count;
	if (defined($mfactor) && $mfactor <= $virtnode_capacity) {
	    $virtnode_capacity = $mfactor;
	}

	#
	# Mark any nodes that are not up with a feature, so that they won't
	# normally get assigned. We want to include them, though, because we
	# allow people to do fix-node to down nodes
	#
	if ($status && ($status ne 'up')) {
	    push @features, "down:1";
	}

	#
	# Mark which site this node belongs to
	#
	if ($site) {
	    push @features, "*&$site:$site_weight";
	}

	#
	# Add any auxiliary types.
	#
	foreach my $auxinfo (@{$auxtypes{$physnode}}) {
	    my ($auxtype,$count) = split(":", $auxinfo);

	    if ($count > $virtnode_capacity) {
		$auxinfo = "$auxtype:$virtnode_capacity";
	    }

	    push(@types, $auxinfo);
	}

	# Add trivial bw spec.
	push @flags, "trivial_bw:400000";

	# Indicate that these nodes are beautiful and unique snowflakes
	push @flags, "unique";

	# Add features
	push(@features, @{$typemap{$ptype}->{'FEATURES'}});
	if (defined($features{$physnode})) {
	    push @features, @{$features{$physnode}};
	}

	my $text = "node $physnode " .
	           join(" ",@types) . " - " . join(" ",@features) .
		   " - " . join(" ",@flags);

	print "$text\n";
    }
}

# Read interfaces
$result = DBQueryFatal("SELECT node_id,card,port,iface,interface_type" .
		     " from interfaces");
while (($node,$card,$port,$iface,$type) = $result->fetchrow_array) {
    $interfacemap{"$node:$card:$port"} = $iface;
    if ((defined $type) && ($type ne "")) {
	$interfacetypes{"$node:$card:$port"} = $type;
    }
}

# Read interface types
$result = DBQueryFatal("SELECT type,max_speed from interface_types");
while (($type,$speed) = $result->fetchrow_array) {
    $interfacespeeds{$type} = $speed;
}

# Read interface switches
$result = DBQueryFatal("SELECT node_id1, iface, node_id2 FROM wires AS w " .
	"LEFT JOIN interfaces as i ON w.node_id1=i.node_id AND w.card1=i.card");
while (($node,$iface,$switch) = $result->fetchrow_array) {
    if ($node && $iface) {
	$interfaceswitches{"$node:$iface"} = $switch;
    }
}

# Read interface cards and ports
$result = DBQueryFatal("SELECT node_id, iface, card, port FROM interfaces");
while (($node,$iface,$card,$port) = $result->fetchrow_array) {
    $interfacecardports{"$node:$iface"} = [$card,$port];
}

$result = DBQueryFatal("SELECT node_id1,card1,port1,node_id2,card2,port2" .
 		    " from wires where type=\"Node\" or type=\"Trunk\"");
while (($node1,$card1,$port1,$node2,$card2,$port2) = 
       $result->fetchrow_array) {
    if ((defined($nodes{$node1}) || defined($switches{$node1})) && 
	(defined($nodes{$node2}) || defined($switches{$node2}))) {
	$iface1 = get_iface($node1,$card1,$port1);
	$iface2 = get_iface($node2,$card2,$port2);
	$iface1bw = get_ifacebw($node1,$card1,$port1);
	$iface2bw = get_ifacebw($node2,$card2,$port2);
	# XXX - This is a bad, bad hack - we  use our knowledge that in the
	# wires table links to the switch always come as node2. We also assume
	# that node interfaces are plugged into switch ports of the same speed.
	# Chris has something better on the way (storing the speed on the
	# switch side), so this is just a temp. hack.
	if (!defined($switches{$node1}) && defined($switches{$node2})) {
	    $bw = $iface1bw;
	} else {
	    if ($iface1bw < $iface2bw) {
		$bw = $iface1bw;
	    } else {
		$bw = $iface2bw;
	    }
	}
	if (defined($switches{$node2})) {
	    $used_switches{$node2} = 1;
	}
	if (defined($switches{$node1})) {
	    $used_switches{$node1} = 1;
	}
	if (defined($switches{$node1}) && defined($switches{$node2})) {
	    # interswitch link
	    if (defined($interconnects{"$node1:$node2"})) {
		$interconnects{"$node1:$node2"} += $bw;
	    } else {
		$interconnects{"$node1:$node2"} = $bw;
	    }
	} else {
	    # !!! - Here we use our knowledge that in the wires table links
	    # to the switch always come as node2.
	    print "link link-$node1:$iface1-$node2:$iface2 $node1:$node1/$iface1" .
		" $node2:$iface2 $bw 0 0 1\n";
	}
    } 
}

#
# If we're supposed to track interswitch bandwidth, subtract out the amount
# that's already in use
#
if ($TRACK_INTERSWITCH_BANDWIDTH) {
    #
    # Get a list of all VLANs
    #
    my $result = DBQueryFatal("SELECT virtual, members FROM vlans");
    while (my ($virtual, $members) = $result->fetchrow()) {
	#
	# Split apart the space-separated list of members
	#
	my @members = split /\s+/, $members;
	my %switches = ();
	foreach my $member (@members) {
	    my ($node,$iface) = split /:/, $member;
	    my $switch = get_ifaceswitch($node,$iface);
	    my ($card, $port) = get_ifacecardport($node,$iface);
	    my $bw = get_ifacebw($node,$card,$port);
	    $switches{$switch} += $bw;
	}

	#
	# Check to see if more than one switch was found among the member
	# list, and if so, go through all the pairs
	#
	my @switches = keys %switches;
	if (@switches > 1) {
	    for (my $i = 0; $i < (@switches -1); $i++) {
		my $switch1 = $switches[$i];
		my $switch2 = $switches[$i+1];
		my $bw = $switches{$switch1};
		if ($switches{$switch2} < $bw) {
			$bw = $switches{$switch2};
		}
		#
		# The trunk link could be listed in either order
		#
		if ($interconnects{"$switch1:$switch2"}) {
		    $interconnects{"$switch1:$switch2"} -= $bw;
		} elsif ($interconnects{"$switch2:$switch1"}) {
		    $interconnects{"$switch2:$switch1"} -= $bw;
		}
	    }
	}
    }
}

foreach $interconnect (keys(%interconnects)) {
    ($src,$dst) = split(":",$interconnect);
    print "link link-$interconnect $src $dst $interconnects{$interconnect} 0 0 1\n";
}

# Print out links
sub get_iface {
    ($node,$card,$port) = @_;
    if (defined($interfacemap{"$node:$card:$port"})) {
	return $interfacemap{"$node:$card:$port"};
    } else {
	# shark hack
	if ($node =~ /^sh/) {
	    return "eth0";
	} else {
	    return "(null)";
	}
	# end shark hack 
    }
};

# Find the bandwidth for an interface
sub get_ifacebw {
    ($node,$card,$port) = @_;
    if (defined($interfacetypes{"$node:$card:$port"})) {
	return $interfacespeeds{$interfacetypes{"$node:$card:$port"}};
    }
    else {
	return $default_switchport_speed;
    }
}

sub get_ifaceswitch {
    ($node,$iface) = @_;
    if (defined($interfaceswitches{"$node:$iface"})) {
	return $interfaceswitches{"$node:$iface"};
    } else {
	return undef;
    }
}

sub get_ifacecardport {
    ($node,$iface) = @_;
    if (defined($interfacecardports{"$node:$iface"})) {
	return @{$interfacecardports{"$node:$iface"}};
    } else {
	return ();
    }
}