GitLab

Probe Cluster. The problem is that the IFB is a shared network that
every node attaches to, which can looks like an ethernet device that
can ifconfig'ed. In other words, one big lan.

But we still want the user to be able to create a lan so that they can
interact with it in thei NS file like any other network.

The NS syntax is:

	set lan2 [$ns make-lan "node1 node2 node3" * 0ms]
	tb-set-switch-fabric $lan2 "infiniband"

The switch fabric tells the backend to do IP assignment for the
specific global network. Yes, I tried to be a little but general
purpose. Lets see how this actually turns out.

This first commit treats the fabric as a single big lan on the same
subnet.

NOTE 1: Since the unroutable IP space is kinda small, but the Probe
Cluster is really big, we can easily run out of bits if we tried to do
assignment on virtual topos. Instead, fabrics get their IP allocation
at swapin time, and the allocations are deleted when the experiment is
swapped out. The rationale is that the number of swapped in
experiments is much much smaller then the number of possible topos
that can be loaded into the DB. Still might run out, but less likely.

The primary impact of above is that IP assignments can change from
one swap to another, but this is easy to deal with if the user is
scripting their experiment; the IP allocation is available via the
XMLRPC interface.

NOTE 2: The current code allocates from a single big network, which
makes it easy for users to mess each other up if they start doing
things by hand. Ultimately, we want each lan in each experinent to use
their own subnet, but that is going to take more work, so lets do it
in the second phase.

The definition of "network fabrics" is in the new network_fabrics
tables. As an example for probe:

	INSERT INTO `network_fabrics` set
		idx=NULL,
		name='ifband',
		created=now(),
		ipalloc=1, ipalloc_onenet=1,
		ipalloc_subnet='192.168.0.0',ipalloc_netmask='255.255.0.0'

missing.

Fix a weird problem where the failure to start up the "crude" listener
got reflected as a failure of the latency test. This was happening on
Fedora 15 because we could not run crude at elevated priority.

Part of the fix is to wait a second after starting a listener and then
make sure it is running. In the case of such an early failure of "crude",
we then try again without elevated priority.

Now accept a range of addresses from first-last, instead of base+mask.
That way, arbitrary ranges can be speficied (address blocks are no
longer limited to aligned powers of two).

We were using "short" and "int" in the data struct passed between client
and server. Note that we still need to address the big/little endian thing.

Hardwire the sort type for a few columns that cause guessType grief.

Remove the insertion of blank-rows-as-spacers in a table. They don't
sort well. Leigh got rid of most of these a while back.

The former ensures that we get the same behavior as when the same
sorted output is produced using "order by" in mysql. I added a
sortable_alpha_cs that we can use for a column should we discover
a case where we really want a case-sensitive sort.

The latter ensures that a user can at least manually do a multi-column
sort (e.g., the equiv of "order by state,eid" when displaying experiments
by clicking on the eid column and then on the state column).

This is needed due to the odd way in which we use sorttable.
In normal use, sorttable.init would be called via the DOMContentLoaded
event hook and this would invoke makeSortable() to do the DOM magic
necessary to ensure all tables marked sortable are make sortable.

But due to our use of AJAX-y dynamic content loading, this won't work
(or so I am told!) So we generate our own call to makeSortable(mytable)
when we produce any sortable table "mytable". But if init() has not been
called before that first call to makeSortable, the DATE_RE regex will
not be initialized and strange things happen in guessType and consequently,
in the sort functions themselves.

Removed now unused variables from both as well.

Imageinfo hides the nastiness of what we need to do to get size or mtime
info about shared images for which the caller does not have direct access.
In these cases we need to make a proxied frisbee master server query and
parse its output.

This allows us to more easily disable a subboss in the event of a temporary
subboss outage (e.g., hardware failure). Previously we would have to remove
the related rows from the DB and restore them later.

Previously ISUP and TBFAILED were the two states that signified that a
node had rebooted to the satisfaction of libreboot (with waitmode==1).
Add RELOAD/RELOADING to that list since the frisbee MFS never sends either
of ISUP or TBFAILED.

Required a modification to TBNodeStateWait() to allow waiting for an
op-mode/state combo as well as just a state.

I made this change in anticipation that it would be useful for more
responsive monitoring of failure in tbswap/os_load. But now I am no so sure.