GitLab

requires the linux threads package to give us kernel level pthreads.

From: Leigh Stoller <stoller@fast.cs.utah.edu>
To: Testbed Operations <testbed-ops@fast.cs.utah.edu>
Cc: Jay Lepreau <lepreau@cs.utah.edu>
Subject: Frisbee Redux
Date: Mon, 7 Jan 2002 12:03:56 -0800

Server:
The server is multithreaded. One thread takes in requests from the
clients, and adds the request to a work queue. The other thread processes
the work queue in fifo order, spitting out the desrired block ranges. A
request is a chunk/block/blockcount tuple, and most of the time the clients
are requesting complete 1MB chunks. The exception of course is when
individual blocks are lost, in which case the clients request just those
subranges.  The server it totally asynchronous; It maintains a list of who
is "connected", but thats just to make sure we can time the server out
after a suitable inactive time. The server really only cares about the work
queue; As long as the queue si non empty, it spits out data.

Client:
The client is also multithreaded. One thread receives data packets and
stuffs them in a chunkbuffer data structure. This thread also request more
data, either to complete chunks with missing blocks, or to request new
chunks. Each client can read ahead up 2 chunks, although with multiple
clients it might actually be much further ahead as it also receives chunks
that other clients requested. I set the number of chunk buffers to 16,
although this is probably unnecessary as I will explain below. The other
thread waits for chunkbuffers to be marked complete, and then invokes the
imagunzip code on that chunk. Meanwhile, the other thread is busily getting
more data and requesting/reading ahread, so that by the time the unzip is
done, there is another chunk to unzip. In practice, the main thread never
goes idle after the first chunk is received; there is always a ready chunk
for it. Perfect overlap of I/O! In order to prevent the clients from
getting overly synchronized (and causing all the clients to wait until the
last client is done!), each client randomizes it block request order. This
why we can retain the original frisbee name; clients end up catching random
blocks flung out from the server until it has all the blocks.

Performance:
The single node speed is about 180 seconds for our current full image.
Frisbee V1 compares at about 210 seconds. The two node speed was 181 and
174 seconds. The amount of CPU used for the two node run ranged from 1% to
4%, typically averaging about 2% while I watched it with "top."

The main problem on the server side is how to keep boss (1GHZ with a Gbit
ethernet) from spitting out packets so fast that 1/2 of them get dropped. I
eventually settled on a static 1ms delay every 64K of packets sent. Nothing
to be proud of, but it works.

As mentioned above, the number of chunk buffers is 16, although only a few
of them are used in practice. The reason is that the network transfer speed
is perhaps 10 times faster than the decompression and raw device write
speed. To know for sure, I would have to figure out the per byte transfer
rate for 350 MBs via network, via the time to decompress and write the
1.2GB of data to the raw disk. With such a big difference, its only
necessary to ensure that you stay 1 or 2 chunks ahead, since you can
request 10 chunks in the time it takes to write one of them.

duplicate this code in the frisbee tree, build a version suitable for
linking in with frisbee. I also modified the FrisbeeRead interface to
pass back pointers instead of copying the data. There is no real
performance benefit that I noticed, but it made me feel better not to
copy 350 MBs of data another time. There is new initialization
function that is called by the frisbee main program to set up a few
things.

of start==end==0! This causes the entire disk to compressed a second time!

pclasses.  This involved removing the heuristics, which, for the most
part, were not worth the cycles they consumed, and scaled badly.

unixgroup_membershit table from the command line. Runs the appropriate
commands to make changes in the 'real world' after the database has been
updated. From the usage message:

Usage: unixgroups <-h | -p | < <-a | -r> uid gid...> >
-h            This message
-p            Print group information
-a uid gid... Add a user to one (or more) groups
-r uid gid... Remove a user from one (or more) groups

a note about suidperl.

so they'll have an address distinct from the 'classic' testbed lists.

config file.

INSTALL_APACHE_CONF in some places and INSTALL_APACHE_CONFIG in
others. Oops.

unchanged to avoid setting up a new testbed-ops list, but this can be changed
if it becomes a problem.

needed to build an emulab boss node in one convenient package.

The config file gets run through autoconf to get the DocumentRoot, log, and
other directories. There is an install target for it, but this is as yet
unused by anything else. There is also a new configure option specifiying where
the config file should go.

standardize on using /usr/bin/perl instead of /usr/local/bin/perl (most already
did.)

messages on tb_compare fails.

from the network, and notes about the ickiness of getting the SNMP MIB
index rebuilt.

This check-in consists of 7 modifications to assign.

1. Equivalence Classes

Defined an equivalence relation on the physical nodes and applied it
to the physical topology to get the resulting quotient topology (abuse
of terminology).  So instead of searching among all possible physical
nodes to make a map, assign only searches among all possible
equivalence classes of nodes.  This tremendously reduces the search
space.  At the time of this writing it reduces the physical topology
from 252 nodes to 13 nodes.  The equivalence classes are generated
automatically from the ptop file.

2. Scoring based on equivalence classes.

Each equivalence class used comes with a significant cost.  This
strongly encourages assign to use equivalence machines when possible.
The result is that an experiment that does not otherwise specify will
almost definitely get machines of the same type.  If this needs to be
reduced in the future it is the SCORE_PCLASS constant.

3. Heuristics

Added a bunch of heuristics for choosing which equivalence class to
use.  This was less successful than I hoped.  A good solution is now
found in record time but it still continues searching.  When OPTIMAL
is turned on these heuristics help a lot.  When off they make little
difference.  I may turn this into a compile time option in the future
since the heuristics do take non-trivial CPU cycles.

4. Fixed the very-very-big-and-evil disconnected-switches bug.

Assign wasn't cleaning up after itself in certain cases.  Disconnected
graphs are now merely a minor, easily ignored, bump rather than the
towering cliffs they use to be.

5. Fixed the not-yet-noticed not-enough-nodes bug.

Found a bug that probably has never come up before because we have
checks that avoid those circumstances.

6. Modified constants.

I was tired of waiting so long for results so, I lowered CYCLES and
reduced the constant for naccepts (Mac, you probably want to add that
inconspicuous number to your configurable constants; look for
"naccepts =").  The results is roughly a speedup of 2.  It works great
currently but we may want to change these numbers up again if we get
problems with features and desires.

7. General clean up.

Associated with the other changes was a lot of restructuring and some
cleanup.  Specifically to the assign loop and scoring code.

it's mostly just a wrapper around snmpit_intel - this is because, when doing
MAC-based VLANs, you never have to talk to more than one switch. When we switch
to doing port-based VLANs with Intels (something we should do soon, but will
take some experimenting), it will likely become a more full-fledged layer.

Not at all debugged yet - won't be able to do that until we have the
mini-testbed more funcitonal.

was originally created.

does not yet exist (but should soon.)

There is a new database table, switch_stack_types, that goes along with these
code changes, to indicate which module should be used for each switch stack.

are created 664, so that group members other than the experiment
creator can swap them.

group directories are now created in a different tree than the
the project directory so that they can be exported independently of
the project tree to the nodes in a group experiment. The tree is
routed at /groups on boss/users and on nodes.

1. mkgroup,rmgroup,mkproj - Minor changes to reflect new group
   directory location (/groups). We leave a symlink in the old spot to
   maintain compatability, and to reduce the number of different
   directories that a person needs to worry about. So, when a group is
   made, you get a real directory /groups/pid/gid, and a symlink
   /proj/pid/groups/gid that points to the former.

2. tmcd/tmcd.c - Minor change to add the additional group directory mount
   in the mounts command. Only done when pid!=gid for the experiment.

3. tmcd/libsetup.pm and friends - Minor changes to fix the fact that
   mkdir does not create subdirs along the way unless the -p option is
   specified. Needed to create the local directory for the mounts
   returned by tmcd for group dirs. Pushed them out to the sup trees,
   although 6.2 images older than the most recent one are not going to
   work right. No one is using those images though, and we should just
   flush the sup trees.

4. exports_setup.in - Ah, the crux of the issue. I really dislike NFS
   at this point. The original idea was to export a third set of
   directories to nodes that were part of a group experiment. Those
   nodes would get /groups/pid/gid exported, and /proj/pid read-only.
   Well, no such luck. On users, /groups and /proj are both really on
   /q, and the old restriction of mountd not allowing an IP to
   specified more than once on the right hand side for any FS, reared
   its ugly head again. As far as mountd is concerned, /q/groups and
   /q/proj are the same thing, and so it bombed when I tried to export
   them on different lines, since that meant an IP was repeated twice.
   So, I reworked exports_setup, and now for any node that is part of
   a group experiment, it gets this:

	/q/proj/pid /q/groups/pid/gid -maproot=root 155.101.132.26

   which at least allows the individual group dirs to be protected
   from each other, but does not allow /proj/pid to be exported read
   only. Sigh.

a good start.

mkprojdir, mkacct-cntrl, mkgroup, and group-update into a set of new
scripts that are more specific to their intended operation, and strive
to do less work.

1. mkacct - Replaces mkacct-cntrl. This script no longer does any
   group stuff. All it does is create new accounts, or update the
   password and gecos fields of existing accounts. Usage is the same
   as it was: "mkacct <userid>", and is typically invoked from the web
   interface via the approveuser form.

2. mkgroup - Replaces group-update. This script creates new groups,
   either for the main project when it is approved, or for subgroup
   creation. This script does not alter the group membership. Usage
   is typically from the web interface, but mkgroup can be invoked
   from the command line: "mkgroup [-b | -a] <pid> <gid>" where -b
   puts it in the background and sends email later, while -a just
   captures the log and emails. This "audit" feature is going to find
   its way into more scripts as soon as I figure out a neat and clean
   perl mechanism to make it easy.

3. setgroups - Replaces group-update. This script modifies the group
   membership of either specific users, or all the users in a
   project. It is typically invoked from the web interface when a
   project leader edits the subgroup membership or when a user is
   first approved to a project or subgroup. Command line usage is:

	setgroups [-b | -a] -p <pid> [user ...]
        setgroups [-b | -a] [user ...]\n

   The first form is mostly a means to speed things up. The web
   interfaces knows exactly what users have need to be changed, but a
   global project update is nice too.

4. mkproj - Replaces mkprojdir. Actually, mkproj still has all that
   directory code, but it also handles creating the groups and the
   account for the project leader. Part of my policy to move as much
   random code out of the web interface and into the PERL backend
   where it belongs.