In principle, almost any research or educational use by those that have a need for it is appropriate and encouraged. This includes use by universities, industrial research labs, and both US and non-US institutions. With some provisos, use for development and evaluation is also acceptable, even by companies. See our posted policies for more detail. If you are unsure about your eligibility to use Netbed/Emulab, please just send us an email inquiry.
If you are new to the Testbed, simply click on the "Start Project" link on the Emulab Home Page. You will need to fill in the forms with your personal information and information about the project. Then click on the "Submit" button. Within a few days you will be contacted via email with an approval message. More information about starting projects can be found in Authorization Page.
If you already have an Emulab account, and wish to start a second project, first log into the Web Interface. Then select the "Start Project" link; all of the personal information will already be filled in. You will need to complete just the project information section.
If you are new to the Testbed, simply click on the "Join Project" link on the Emulab Home Page. You will need to fill in the form with your personal information, and provide the name of the project you are trying to join (typically, the Project Leader will have told you the name of the project). Then click on the "Submit" button, and wait for an email with your new user key. When that email arrives, use the link in it (or the key itself), and use it with your password to log into the web site and verify your account. Then just wait for the project leader to approve you. When approved you will receive an email message saying so, and you can then log into the Testbed.
Joining a project has 3 stages. The first two are done by the person trying to join, and they both must be completed before you can approve their application. The first two are outlined in the previous question, where the user fills out the "Join Project" form, and performs account verification. After these two steps are both complete, the project leader and any group leaders in the group (More info here) will get an email saying the account is ready to be approved, and it will appear on the list of new users waiting to be approved.
If someone says they've applied, but you haven't received an email from Emulab about it, and they don't show up on your list, the most likely cause is that they haven't finished the verification step.
Yes! Emulab uses email notifications to you in several different ways. Often it will send you a copy of information regarding experiments you set up, applications to projects, and other things you do at Emulab. Sometimes (like with account verification) the email is a critical part of being able to use Emulab.
For those reasons it is critical that any spam filtering software you have accept email from Emulab itself (anything coming from the emulab.net domain) and from Emulab staff (from the cs.utah.edu or flux.utah.edu domains). In many cases, it may also say that is from a specific machine in those domains. Our messages usually do not get flagged as spam by most filters, but in certain cases it can be a problem. It is also important that it not require manual intervention or confirmation to get emails through to you, so programs like SpamKiller can cause problems.
It is also important to read your email often while you are using Emulab, especially while you have machines reserved in an experiment. A few emails may be the only notification you'll get before we swap out an experiment that appears to be idle, and if you don't respond, you may lose important work. Email is also our method for informing you about problems we may be experiencing, downtimes, or other important announcements. Your experience with Emulab (and ours with you) will be much more pleasant if everyone is responsive to email.
Once you have been approved to start (or join) your first project, you will be able to log into Emulab's user machine, users.emulab.net. We require that all Emulab users use ssh. For example, if your Emulab account name is "joe", then you would do:
ssh users.emulab.net -l joe
Your password is the same as the password you supplied to the Start (or Join) Project web page.
There are several useful (although not required) programs installed on users.emulab.net in /usr/testbed/bin. You should edit your dot files to include that directory in your search path.
Yes. You may join (and/or start) as many projects as you like, subject to Emulab administrative policies.
Yes. You can change your Emulab Web password and your Emulab login password (the password you use to log into users.emulab.net, as well as nodes in your experiments). To change your password, simply click on the "Update User Information" in the menu to your left, and then enter your new password in the location provided. Your new password will be installed on users.emulab.net immediately. Your experimental nodes will get the new password when they reboot.
Yes. To designate a TA, you must first create a project group. A project group is a lot like a unix group, and in fact unix groups is the mechanism used to protect members of one group from members of another group. When you create a group, you designate a group leader who is responsible for approving users who apply to join the group. Group leaders may also terminate experiments that have been created by members of the group. As Project Leader, you may also shift members of your project in and out of your project's groups as you like, and you are automatically a member of all groups within your project. As a convenience, all new projects are created with one new group, termed the default group. As its name implies, whenever the group is left unspecified in a form, it defaults to the project group (this allows you to create a project without any sub groups at all; new members join the default group, new experiments are created in the default group, etc.).
Project groups are created via the Project Information link at your left. Simply go to the project page in which you want to create a group, and look for the "Create New Group" link. More information on project groups is available via the Emulab Documentation page in the Groups Tutorial.
Don't hesitate to send us email!
Ok, hesitate just a little and read the rest of this entry first. Before sending email, be sure to check out the Troubleshooting entry which describes several common problems and possible causes. If you do send email, there are several pieces of information you should include to make our job easier.
If you cannot find an answer to your question in the Emulab Documentation, then you can send us an email message. We will try to answer your question as quickly as we can.
Yes, we have an extensive tutorial on using the Testbed.
Yes, we provide a GUI that gives you an easy to use drawing palette on which you can place nodes, lans, and links. Testbed specific attributes such as operating system, hardware type, and link/lan characteristics, may be attached to each object. With a single click, you can instantiate your new topology on the Testbed as an experiment in one of your projects. Alternatively, you can save the auto-generated NS file on your machine, edit as required, and then submit it later when creating an experiment.
To access the GUI, please log in and go to the Begin Experiment page. Note: you need a Java compliant browser.
Yes, but only those imposed by the physical hardware that is currently available in our testbed. The constraints that people most commonly run into are the maximum speed of our links (100Mbps) and the maximum number of network interface cards (NICs) in our machines. You can't get any links faster than 100Mbps, since we don't yet have gigabit links for experimenters to use. Our nodes each have 4 experimental network interfaces, so each node can be a member of up to 4 links or LANs. A good strategy for making your topology fit within those limits is to replace multiple links to a node with a LAN or with a router node.
Another approach is to use emulated network links. Emulated links are multiplexed over the actual physical links, up to the physical bandwidth that the physical link can support. In other words, five 20Mb links can be multiplexed over a 100Mb physical link. More information on emulated links can be found here. Ask Testbed Operations if you need further assistance.
You can ask for as many nodes as are currently available! You can click on the "Node Reservation Status" link at your left to see how many nodes are currently free. If you ask for more than are currently available, your experiment will be rejected (you will receive email notification shortly after you submit your NS file to the web interface).
We urge all new Emulab users to begin with a small 3-4 node experiment so that you will become familiar with NS syntax and the practical aspects of Emulab operation.
You can keep them as long as you need them, subject to our Node Usage Policies. In general, you should do your work, and then terminate your experiment as soon as you're done with it. If you're not done with it, but are through for a while, you should probably "swap out" your experiment (See the question What is Swapping in this FAQ). It is especially important to swap out your experiment if you're through with it for the weekend. Emulab usually gets heavy use on the weekends by users who need to make very large experiments, so it is important to leave as many nodes available as possible.
It is better for you and us if you don't just keep hitting the submit button every few minutes. It wastes your time and floods us with email (we get every failure message you do!) Instead, you can now use the Batch System to queue an interactive job. By submitting your experiment as a batch job, but without any tb-set-node-startcmd directives in your NS file, the job will be queued until nodes are available. For most experiments, this means just using your regular NS file, and checking the Batch Mode Experiment box when you create the experiment.
When your queued job is swapped in, you will be sent email to inform you, and you can start working! Please note that the experiment will be idle when it is swapped in, and will be idle swapped if you do not get things running on the nodes in a short period of time. If your experiment does get swapped out before you can get to it, you can always visit the experiment's information page and try again by using the Queue Batch Experiment menu item.
For example, say you need 50 nodes but there are only 40 free. In general, getting this many nodes is going to require intervention from Testbed Operations, if only so we can ask other experimenters to free up nodes, if possible. Please send us email if you are not able to able to get the number of nodes you need for your experiment.
Another alternative is to use the Batch System. If your experiment is amenable to being batched (does not require human intervention to start and stop), then you can submit a batch request, which will be serviced when enough nodes become available. Typically, you would start out with a few nodes, getting used to the batch system and creating whatever scripts are needed to make the experiment batchable. Then scale up to larger numbers of nodes. Thats the easiest way of getting a lot of nodes!
Yes. Project leaders get root access to all of the nodes in all of
the experiments that are running in their project. Project members
get root if their project leader grants them root access, when the
leader approves the group membership request.
Root privileges are granted via the sudo
command. The
tutorial describes
this in more detail.
Yes. Each of the PCs has its own serial console line with which you can interact, either directly from your desktop (see next FAQ entry), or by hopping through the "users" machine, using our console program. To connect over serial line to "pc1" in your experiment, ssh into users.emulab.net, and then type console pc1 at the Unix prompt. You may then interact with the serial console (hit "enter" to elicit output from the target machine).
In any case, all console output from each node is saved so that you may look at it it later. For each node, the console log is stored as /var/log/tiplogs/pcXXX.run. This run file is created when nodes are first allocated to an experiment, and the Unix permissions of the run files permit only members of the project to view them. When the nodes are deallocated, the run files are cleared, so if you want to save them, you must do so before terminating the experiment.
Clicking "Connect to Serial Line"
in the Node Options page will send your browser a "text/x-testbed-acl"
".tbacl" file.
In windows, if you have installed tiptunnel
, available
below, you can save this file in a folder and double-click it
to launch a tunneled connection to your node.
In FreeBSD or Linux, you can save the file and pass it as an argument
to tiptunnel
, or associate it with
tiptunnel
in your web browser.
Upon connection you typically first have to hit "enter" to
elicit output from the target machine.
tiptunnel
installer for Windows here.
tiptunnel
statically-linked x86
binary for FreeBSD here.
tiptunnel
statically-linked x86
binary for Linux here.
gunzip
,
then tar xvf
on the downloaded file.tiptunnel
binary into
a directory of your choice (/usr/local/bin
,
or ~/bin
are two good places.)tiptunnel
,
a new xterm window with a telnet session open to your node
should emerge.tiptunnel
directly;
this may be more convenient than using the
web interface every time you wish
to connect to a node in your experiment.
Note that these files are valid for the
lifetime of your experiment.)preferences
from the edit
menu.Navigator
, then Applications
under
it.New...
button.MIMEType
box, type text/x-testbed-acl
Suffixes
box, type tbacl
Application
in the Handled by
boxApplication
, either type the path to the
tiptunnel
binary, or use Choose...
to find
it.%s
after the
path to the application in the box. This tells netscape to actually
pass the aclfile into tiptunnel (Mozilla does not require this;
see below.)OK
, then OK
again.preferences
from the edit
menu.Navigator
, then Helper Applications
under it.New Type...
button.MIMEType
box,
type text/x-testbed-acl
File extension
box, type tbacl
Application to use
, either type the path to the
tiptunnel
binary, or use Choose...
to find
it.%s
.OK
, then OK
again.Yes. Each of the PCs is independently power controlled. If your node hangs, or is otherwise unresponsive, you can use the node_reboot command, as discussed in the Emulab Tutorial.
If you manage to corrupt a disk (or slice), no worries. You can easily repair the damage yourself by reloading a fresh copy of the default disk image. You will of course lose anything you have stored on that disk; it is a good idea to store only data that can be easily recreated, or else store it in your project directory in /proj. Disk reloading is covered in more detail in the Emulab Tutorial.
Each project has its own directory, rooted at /proj, which is available via NFS to all of the nodes in experiments running in that project. For example, when the "RON" project was created, a directory called /proj/RON was also created. This directory is owned by the project creator, and is in the unix group "RON." Its permission (mode) is 770; read/write/execute permitted by the project creator and by all of the members of the project RON, but protected against all access by people outside the RON project.
Sub-groups within a project likewise have a directory in the /groups tree. A group named "group1" in the RON project would thus have a group directory in /groups/RON/group1.
Project members are encouraged to store any files needed by their experiments in the corresponding /proj or /groups directory.
Yes. All of the files in your home directory on /users, all of the files in your project directory in /proj, and all of the files in your groups directory in /groups are filesaved. While we can restore lost files in an emergency, we encourage you to back up critical data on your own to avoid (possibly long) delays in conducting your experiments.
No! The nodes in your experiment are not filesaved. Any changes
you make to the local filesystems will be lost if the event of a
disk failure. We plan to provide a mechanism for experimenters to
create snapshots of their node state, but that is not done yet. In
the meantime, any files that must not be lost should be stored in
the project directory (/proj/
Swapping is when you (or we, or the Emulab system) temporarily swaps out your experiment, releasing all of the nodes in the experiment. Your experiment is still resident in the Emulab database, and you can see its status in the web interface, but no nodes are allocated. Once an experiment is swapped out, you can swap it back in via the web interface by going to the Experiment Information page for your experiment, and clicking on the swapin option. You can also modify it.
The idle-swap checkbox in the Begin Experiment web page is used to determine what experiments can be automatically swapped by the testbed scheduling system. Note that all experiments are capable of being swapped; even if you do not check the idle-swap box, you are free to swap your own experiments as you like. The only difference is that the testbed scheduling system will not consider your experiment when looking for experiments to swap out. You will sometimes notice that the Experiment Information page does not contain the swap link. That is because experiments cannot be swapped when they are in transition. For example, when the experiment is being swapped in (say, after first being created) the link will disappear until the experiment is fully swapped in, and it is capable of being swapped out. You will need to occasionally reload the page so that the updated state is recognized and the swap link appears.
Be aware that we do not currently save any files that you may have placed on your nodes. When your experiment is swapped back in, you will likely get different nodes, with fresh copies of the disk images. For that reason, you should not swap your experiment out unless you make arrangements to save and restore any state you need.
Please be sure to read our Node Usage Policies, which contain detailed information on swapping.
Experiment restart (or perhaps more aptly, replay) allows you to rerun your experiment from scratch, but without the added expense of a swapin and swapout. In other words, the nodes that are currently allocated to your experiment are all rebooted, and the experiment startup state is cleared. The event scheduler for the experiment is restarted, and your event sequence is replayed again. Note that your rpms and tarfiles are not installed again. Replay is obviously faster than swapout/swapin, and has the added benefit that you will not run the risk of not being able to swapin for lack of available nodes.
We have a command called portstats
that allows you access
to some of the port counters on our switches. To use it, you'll need
to ssh to users.emulab.net. 'portstats <proj>
<exp>
' will get you stats for all experimental interfaces in
your experiment. Run 'portstats -h
' to get a list of other
options, such as different sets of stats.
Note that the numbers returned by portstats
do not get
reset between experiments.
We set up names for your nodes in DNS, for use from outside,
and /etc/hosts
files for use on the nodes in the experiment.
Since our nodes have multiple interfaces (the control network, and,
depending on the experiment, possibly several experimental interfaces,)
determining which name refers to which interface can be somewhat
confusing. The rules below should help you figure this out.
node.expt.proj.emulab.net
in DNS,
so that they visible anywhere on the Internet. This name always refers
to the node's control network interface, which is the only one
reachable from the outside world.
/etc/hosts
file.
node.expt.proj.emulab.net
. (note
that, since we put .emulab.net
in nodes' domain
search paths, you can use
node.expt.proj
as a shorthand.)
This name always refers to the control network
node-link
form - You can refer to an
individual experimental interface by suffixing it with the name of
the link or LAN (as defined in your NS file) that it is a member
of. For example, nodeA-link0
or
server-serverLAN
. This is the preferred way to refer
to experimental interfaces, since it uniquely and unambiguously
identifies an interface.
nodeA
.) Note that this differs from the fully-
qualified name in that no domain is given. We also create short
names for nodes you are not directly connected to.
However, if two nodes are
connected with more than one interface, or there is more than
one route between them, there is no guarantee that the short name
has been associated with the one is on the 'best' (ie. shortest or
highest bandwidth) path - so, if there is ambiguity, we strongly
suggest that you use the node-link
form.
Yes. On the experiment view page, choose "Modify this Experiment". This will allow you to modify an experiment, either swapped-out or in, by editing its NS file.
If the experiment is swapped-out, Experiment Modify will simply replace its topology with the newly specified one; this new topology will be mapped when the experiment is swapped in.
If the experiment is already swapped-in, Modify will change the topology and map in the portions which have been changed. This allows dynamic addition, subtraction, and replacement of an experiment's nodes and links. However, when modifying swapped-in experiments, there are a couple things to keep in mind:
eventsys_control [-f] replay proj expt
Yes. We provide sources and packages for a few versions of FreeBSD and RedHat Linux. The place to look for available software is under /share on either users.emulab.net or your experimental nodes. This path is readonly (and NFS mounted on the nodes), so you'll need to make a copy of the contents found there if you need to do more than reference them for information or installation.
Please see the Hardware Overview page for a description and count of the computers that comprise the Testbed.
See the previous FAQ entry for a description of what hardware is available.
If you would like to ask for a specific type of hardware, such as a
pc850
, see the
tb-set-hardware
command in our NS
extensions document. You can also define classes of nodes which should
be given the same hardware using
virtual
type commands.
If you click on the "Node Reservation Status" link in the menu to your left, you will see a summary of the number of nodes (by type) that are currently available, followed by a listing of the reservation status of each individual node.
Yes! You can specify the delay, bandwidth, and packet loss rate between any two nodes in your topology. Bandwidth and delay are specified in the NS duplex-link statement, while packet loss rate is specified with the Emulab tb-set-link-loss extension to NS. You may also specify delay, bandwidth, and packet loss rate between nodes in a regular LAN.
Please see the Extensions page for a summary of all Emulab NS extensions, and the Emulab Tutorial for an example.
Yes! If your NS file specified traffic shaping on a link, then you can subsequently modify those parameters after the experiment has been swapped in. Note that you cannot convert a non shaped link into a shaped link; you can only modify the traffic shaping parameters of a link that is already being shaped. To modify the parameters, go to the Experiment Information page of your experiment, and click on the "Control Traffic Shaping" menu option. Follow the instructions at the top of the page.
An alternative method is to log into users.emulab.net and use the delay_config program. This program requires that you know the symbolic names of the individual links. This information is available via the web interface on the Experiment Information page. The command line syntax for delay_config will be displayed when the -h option is given.
Yes! Please see the advanced tutorial. Note though, that these other parameters can be specified for duplex links only (not lans), and that they are not configurable with delay_config, but with a different testbed utility call tevc (also described in the advanced tutorial).
Please see the Software Overview page for a description of the Operating Systems that can be run on each of the Testbed nodes.
When a choice of OS is available, you may specify which one you prefer for each node in the NS file using the Emulab tb-set-node-os extension to NS. When your experiment is configured, the appropriate disk image will be loaded on your nodes, and the selected operating system will boot up on each.
Please see the Extensions page for a summary of all Emulab NS extensions, and the Emulab Tutorial for an example.
Yes! If you have an RPM or Tarball (or more than one) that is appropriate for loading on the OS you have selected, you can arrange to have them loaded automatically when your experiment is configured. The Emulab NS extension, tb-set-node-rpms, is used in the NS file to specify a list of RPMS to install. The tb-set-node-tarfiles command is used to specify a list of Tar files to install, as well as the directories in which the tars should be unpacked. You may specify a different list for each node in the experiment. The RPMs/Tarballs will be installed when the nodes first boot after the experiment is swapped in or if the nodes detect a changed RPM/Tarball during a reboot.
Please see the Extensions page for a summary of all Emulab NS extensions, and the Emulab Tutorial for an example.
Yes! You can arrange to run a single program or script when your node boots. The script is run as the UID of the experiment creator, and is run after all other node configuration (including RPM installation) has completed. The exit status of the script (or program) is reported back and is made available for you to view in Experiment Information link in the menu at your left. The Emulab NS extension tb-set-node-startcmd is used in the NS file to specify the path of the script (or program) to run. You may specify a different program for each node in the experiment.
Please see the Extensions page for a summary of all Emulab NS extensions, and the Emulab Tutorial for an example.
By default, we do not setup any static routes or run any routing daemon on nodes in an experiment. However, we do provide several options for experimenters, which are described in the "Setting up IP routing between nodes" section of the Emulab Tutorial.
If your application requires synchronization amongst your nodes, you may use the Emulab provided synchronization server, which provides a very simple form of barrier synchronization. Use of the synchronization server is described in more detail in the Emulab Tutorial.
Yes! You can run your own OS (or a customized version of an Emulab supported OS) on any of the PCs. You can also run OSKit kernels on the PCs. Each of the PCs is partitioned with two DOS partitions large enough to hold the typical OS installation. The 1st and 2nd partitions are each 3GB. The 3rd partition is 500MB, and is labeled as Linux Swap. The 4th partition is the remainder of the disk, and varies in size depending on the pc type. We recommend that you use the 1st or 2nd partition; using the 4th partition will restrict the number of machines that you can run your OS on since it varies in size. Note that you must leave the MBR (Master Boot Record) in sector 0 alone, and that your custom partition must contain a proper DOS boot record in the first sector.
Please note that while users are free to customize their disks and install their own operating systems, Emulab staff will not be able to offer more than encouragement and advice! We cannot install the OS for you, and we cannot load CDROMS, floppy disks, or tape drives! We do provide an easy way for you to boot FreeBSD from a memory based filesystem (MFS) so that you can more easily work with the disk (in case it is not possible to install your OS on a live disk). Beyond that, you are pretty much on your own!
Many users had great success with customizing an Emulab supported OS (FreeBSD or Linux), and then creating a disk image that is autoloaded when the experiment is swapped in. We strongly encourage people to use this approach whenever possible! There is more information available in the Custom OS section of the Emulab Tutorial.
No. At this time you cannot share OS images between projects. We are thinking of adding project collaboration support, but that is a future project.
In the meantime, you will need to create an image descriptor in the project that wants to use your image. Fill out the form, but leave out the "Node to Obtain Snapshot from". Then just copy the image over to the default path it picked for you in the form. There is more information available in the Custom OS section of the Emulab Tutorial.
Each node has a partition at the end of the disk that you can use if
you wish. In Linux, the partition is /dev/hda4
; in FreeBSD,
it's /dev/ad0s4 . There is no filesystem on this partition,
so you'll need to create it yourself. Before going any further, there
is one very important point: anything you put in this disk space will
be lost when your experiment swaps out! That is, unless you create
a Custom disk image
before it gets swapped.
Recent versions of our standard FreeBSD and Linux images include a
script, /usr/testbed/bin/mkextrafs
for this purpose.
Just do the following as root:
/usr/testbed/bin/mkextrafs /mnt
and it will create the filesystem, mount it on /mnt
and make an entry in /etc/fstab
so that the filesystem
will be mounted on future reboots.
If that script does not exist, you can perform the steps by hand. For example, in Linux do the following as root:
fdisk /dev/hda
mkfs /dev/hda4
mount /dev/hda4 /mnt
You may want to add the filesystem to /etc/fstab
so that it will be automatically mounted on future reboots.
fdisk -i4 /dev/ad0
disklabel -w ad0s4 auto
newfs /dev/ad0s4c
mount /dev/ad0s4c /some/where
You may want to add the filesystem to /etc/fstab
so that it will be automatically mounted on future reboots.
The available space ranges from 6-33GB depending on the disk type.
Probably not. By default, the testbed startup scripts currently start two daemons in addition to the OS's standard set. Other daemons may be started depending on the network services you ask for in your ns file (see below).
Unconditionally started daemons:
healthd
- A low overhead hardware health monitor.
This deamon periodically polls the machine's health monitoring
hardware and sends this information back to our boss
node for analysis. The hardware is polled once per second, and a
status datagram is sent out once every five minutes.
Healthd
's overhead is quite low, but it can be safely
killed and disabled from startup if you're worried about possible
side effects. It is started by
/etc/testbed/rc.healthd
.
slothd
- A low overhead usage analysis tool.
Slothd
is important to efficient testbed utilization
and should run on every node whenever possible. Its overhead is
almost negligible (essentially less than running 'ls -l
/dev'
once every five minutes), and should not interfere with your
work. However if your experiment is exceptionally sensitive, then
you may arrange with us to disable slothd
. Please
note that we will restart this daemon if it is not running unless
prior arrangements have been made.
Conditionally started daemons:
gated
- A network routing daemon.
If you have requested automatic routing on your nodes with
$ns rtproto Session
in your NS file, this will
start gated
on all of your nodes.
We have left all daemons started by the operating systems' default
configurations (such as cron
) enabled, so you should also
look at them if you are concered about running processes affecting
your experiment.
In short, yes, the local nodes in Emulab (but not all remote
Netbed nodes) support IP Multicast on the experimental
network. In order to use it, you must have a kernel that
supports it, and if you want multicast routing, you'll need to
enable mrouted
. (You can do it manually, or
automatically via program objects or startup commands, but the
rtproto commands will not do it.)
Yes. Emulab blocks all of the low numbered ports (ports below 1024), with the exception of ports 20 and 21 (FTP), 22 (Secure Shell), and 80 (HTTP). This is for the protection of experimenters, as well as to ensure that an errant application cannot become the source of a Denial of Service attack to sites outside of Emulab. If your application requires external access to other low numbered ports, please contact us to make special arrangements.
Experiments can fail in many, many ways, but before you send the above vague question off to us, consider a couple of things. First, look carefully at the "experiment failed" e-mail that you received. It includes a log of the setup process which, while not a model of clarity, often contains an obvious indication of what happened.
One potential point of failure is the mapping phase where Emulab
attempts to map your topology to the available resources. Look
in the log for where it runs assign
. Common errors
here include:
*** NN nodes of type XX requested, but only MM foundin the log. You should always check the free node count on the left menu before trying an experiment swapin. Keep in mind that shaped links might require additional traffic-shaping nodes above and beyond nodes that are explicit in your topology.
If the setup log shows assign
failing repeatedly and
eventually giving up, contact us.
The next potential failure point is the setup of the physical
nodes. If you are explicitly setting the OS image to use with
tb-set-node-os
, then make sure you have specified
a valid image (e.g., did you spell the OS identifier correctly?)
Again, the log output should include an error if the OSID was
invalid. Try:
os_load -lon users.emulab.net to get a list of OSIDs that you can use.
If the OSID is correct, but the log contains messages of the form:
*** Giving up on pcXXX - it's been NN minute(s). *** WARNING: pcXXX may be down. This has been reported to testbed-ops.then a node failed to reach the point where it would report a successful setup to Emulab.
Near the end of the experiment setup, Emulab's event system can fail to startup with a message like this:
Starting the event system. *** ~/.ssh/identity is not a passphrase-less key You will need to regenerate the key manually *** /usr/testbed/devel/stack/sbin/eventsys.proxy: Failed to start event system for foo/barOr, like this:
Starting the event system. Permission denied, please try again. Permission denied, please try again. Permission denied. *** /usr/testbed/devel/stack/sbin/eventsys.proxy: Failed to start event system for foo/barThis failure occurs because you have manually changed your default SSH identity (~/.ssh/identity) or edited your authorized_keys file in your Emulab home directory without going through the "Edit SSH Keys" web form on your user page. The easiest way to fix this is to make sure the passphrase is empty using ssh-keygen(1) on the user's machine:
users$ ssh-keygen -p -P "<old-passphrase>" -N "" -f ~/.ssh/identityThen, make sure the corresponding public key in your Emulab home directory ("~/.ssh/identity.pub") is listed in the "Edit SSH Keys" form.
Such failures can be caused by many things. Sometimes
a transient load on an Emulab server can push a node over its
timeout, though this is happening less and less as we
improve our infrastructure. Most often, these failures are caused
by the use of custom images which either do not boot or do not
self-configure properly. These are harder to dianose because you
often need access to the console logs to see what happened,
and these logs aren't available after an experiment fails.
However, it is possible to interactively monitor
the console while the experiment is setting up since console access
is granted early in setup process. You can either use the
console
command on users, use the
tiptunnel client application,
or just run "tail -f" on the /var/log/tiplogs/pcXXX.run
file.
The most common reason is that your topology includes nodes which are not directly connected, and you have not setup any routing. Refer to "How can I turn on routing or set up routes automatically in my nodes?" for details. If you cannot send packets between two machines which are directly connected (via a link or a lan), then there are two possibilities: either the nodes did not properly negotiate their speed and duplex with the Cisco switch, or the physical wire is loose or bad. In these cases, you should contact us for help.
The most likely problem is that it is using the unshaped control network for the traffic you're looking at. This occurs when it tries to contact a node using a "pcXXX" address, like pc76 or pc76.emulab.net, or when it tries to ping a fully-qualified name, like NodeA.myexpt.myproj.emulab.net, which also resolves to a control network address. On one of your nodes, take a look at the file /etc/hosts. It shows the IP addresses and aliases that refer to the different experimental interfaces. These are the names/IPs you can use to see the delays.
See this section of the tutorial for more details on the control network. For a discussion of the way to 'name' interfaces on the control and experimental networks, see the the naming section of this document.
You are probably pinging through the control net interface. See this Troubleshooting FAQ entry and the control net section of the tutorial.
Short answer: Ping is round trip, PLR and delay are "one way".
Long Answer: If you're not seeing any traffic shaping at all (100Mbps, 0ms, 0plr), see this FAQ entry. If you are seeing shaping, but something different than you expected, it is probably because link characteristics are one way, and you're measuring them over the round trip.
For instance, if you asked for a link that was 100Mbps, 30ms, with 5% (0.05) packet loss rate (plr), you may expect ping to show 30ms ping times and 5% loss rate. But what you should see is 60ms latency for the round trip, and a loss rate of 9.75%. Latencies can be added, therefore 30ms + 30ms gives 60ms. However, loss rates are probabilities, and must be multiplied. The chance of a packet making it across a 5% lossy link is 95%, so with a 95% chance of arriving at the destination, and a 95% chance of returning if it made it there, and the total chance of making a round trip is .95 * .95 = .9025 or 90.25%, or a round trip loss rate of 9.75% on a 5% lossy link.
Short answer: Decreasing the bandwidth of a link means that your bytes take longer to get where they are going!
Long Answer: A ping packet is 98 bytes of data; 56 bytes of data plus 8 bytes of ICMP header plus 20 bytes of IP header plus 14 bytes of ethernet header. At 100Mbs those 98 bytes takes .0078ms to traverse the wire, which is hardly noticeable! If you have set the delay of your link to 10ms, then your ping packets will incur 10ms+0.0078ms of delay in each direction, for a 20ms roundtrip time.
Say you set the bandwith of your link to 250Kbs. The wire time for those same 98 bytes is now 3ms. If your delay is 10ms like above, then your ping packets will incur 10ms+3ms of delay in each direction, for a 26ms roundtrip time! If you set the bandwith to 100Kbs, the wire time is now 7.8ms and your ping packets will incur 10ms+7.8ms of delay in each direction, for a 35.6ms roundtrip time!
Note: If you have a router connecting two nodes, then each of the two links will incur the same wire time (and delay of course). In the above 250Kbs example, each ping packet will incur 3ms of wire time to the router and another 3ms from the router to the destination. The ping reply packet will see the same 6ms of wire time. If your delay is again 10ms, then the ping roundtrip is 52ms.
Short answer: TCP needs large send and receive socket buffers in order for its throughput to approach the capacity of long fat networks (LFN) i.e. link/LANs with a large bandwidth-delay product (BDP). Use UDP instead, if your intention is just to test the bandwidth. If you need to tune the throughput of your TCP application, refer to "Enabling High Performance Data Transfers".
Long Answer: In order to observe the bandwidth that you specify, it is necessary to keep the data pipe between the sender and the receiver full. For a reliable window based protocol such as TCP, the window size represents the number of unacknowledged bytes. TCP needs to keep the unacknowledged bytes around until the acks for them are received. These bytes are retained in socket buffers. On a link/LAN with bandwidth B and round-trip-time (RTT) D, the sender TCP needs to be able to transmit B times D bytes before expecting any acknowledgement, if the data pipe has to be kept full. The effective sending window is dependent on receiver advertised window besides other things. It is necessary to have sender and receiver socket buffers at least as high as BxD. If you only care about optimal end-to-end TCP throughput, then the socket buffers need to be BxD where B is the bandwidth of the bottleneck link and D is the end-to-end RTT. Refer to the following "TCP mini-tutorial" or a basic TCP/IP book for the gory details.
Since UDP is unreliable and not flow controlled, you may just be trying to push packets out on the wire too fast. In this case packets will be dropped before even getting to the wire. You may also be dropping packets at the receiver side if the consumer cannot keep up.
If you are trying to saturate a 100Mb link and can't do it, it may be due to using too small a packet size. For small packet sizes, the limitation on a 100Mb link will be the packet rate, not the raw bandwidth. The eepro100 ethernet NICs in most of our machines can only generate a little over 100,000 packets per second (pps). With 64-byte packets, you will thus only see about 51.2Mb/sec of raw data or about 14.4Mb/sec of UDP payload (a 64 byte ethernet packet can hold only 18 bytes of UDP payload). Note that the theoretical max is less than 150,000 pps for 64-byte packets, so even better NICs would not change this.
Considering that stock FreeBSD and Linux can generate even fewer packets per second due to interrupt and scheduling overheads, you probably need to use at least 200-byte packets to saturate a 100Mb link.
The routing daemons are probably talking to each other via the control net and routing traffic through it since it is the shortest path. You will need to configure your daemon to ignore the control net interface. See the control net section of the tutorial.