Compute clusters for Debian development and building final report
Summary
-
The goal of this project was intended to have Eucalyptus cloud to
support ARM images so it allows Debian developers and users to be able
to use such facility for taks such package building, software
development (ie. Android) under a Debian pre-set image,
software testing many others.
What was expected to have at the end is a modified version of Eucalyptus
Cloud that supports ARM images on the first place. To date this goal has
been reached but is not complete, read production-ready. Extensive test
needs to be done. Besides that we have another goal which is to get
Debian community to use this new extended tool.
Project overview
-
Eucalyptus is a hybrid cloud computing platform which has a Open Source
(FLOSS) version. It s targetted to be used for PaaS (Platform as a
Service), IaaS (Infrastructure as a Service) and other distribution
models. It can be used also for cloud management. Given that it has
implemented the EC2, for computing, and the S3, for storage, API, it s
compatible with existing public clouds such Amazon EC2. Currently it
supports running i386 and amd64 images or NC (Node Controllers) in
Eucalyptus naming.
Eucalyptus is a complex piece of software. It s architechture it s
modular composed by five components. Cloud Controller (CLC), Walrus (W),
Storage Controller (SC), Cluster Controller (CC) and Node Controller
(NC). The first three are written in Java and the remaining are written
in C. Our project modifications was targetted to the NC component,
altough there is a remaining task for hacking the UI to allow setting
the arch of the uploaded NC image instance.
The Node Controller is in charge of setting the proper information for
Eucalyptus to be able to run virtualized images. Eucalyptus uses XEN and
KVM hyphervisor to handle internal virtualization subsystem, and
Libvirt, which is an abstraction of the existing virtualization
libraries, for program interfacing.
Having that in mind we are back to our project. For our project to be
sucessful we had various tasks to do. Beginning with understanding such
complex piece of software, followed by hacking the requiered bits and
later integrate the work so it results in a useful tool. Our approach to
the project was then inline with such description.
The project s history
I started my participation in GSoC approaching Steffen to discuss about
what was expected, or more accurately, what was in his mind. We
exchanged emails previous to my application and that resulted on my
application being submitted. From the beginning I should tell that I
wasn t much clear on the final product of the project, however we
refined ideas and goals during the weeks following to the official
begining of the project. What was clear for me, after all, is that for
this to see the light and gain adoption Eucalyptus code needed to be
dealt with.
Our first task was review ARM image creation from scratch, using the
work done by Aurelien and Dominique in previous GSoC. I ve managed to
get an updated ARM image running under qemu-system-arm. The issues
presented in the past were almost unexistant now that the versatile
kernel is official. You can see this on my first report.
After that being done and sometimes in parallel my main goal was then understand
the internals of the Eucalyptus software in order to figure out it s
feasibility, where do we need to extend and how big the task is, we
didn t knew upfront. From the beginning of the project Steffen was kind
to introduce me to the Eucalyptus developers and that have resulted in a
good outcome for Debian, IMHO, to date.
Understanding Eucalyptus internals was quite a fun task to say the less. As you
can see on the first part of the report Eucalyptus is modularized and
the component I was expected to work with was the NC (Node
Controller). Given that I isolated my work focus, I started to focus on
learning the internals about that component.
Node Controller as described is in charge of talking with
virtualization hyphervisors in order to arrange things to run guest
image instances. The module has basically one main component: handler.c,
in charge to talk to the appropiate hyphervisor and there are
extensions (think of OOP polymorphism but acknoledge it s plain C)
that interact with KVM or XEN.
I figured that if qemu-kvm is ran in Hyphervisor mode we can manage to
run an ARM image with qemu-system-arm. Given that Eucalyptus has
interaction with KVM hyphervisor in place, this answered the question of
the project s feasibility. First green light on.
>From this point the scope reduced to interacting with the KVM handler
(node/handler_kvm.c) and extend it to support running an image that is
not amd64 or i386.
NC makes use of libvirt to abstract interaction with Hyphervisors. So,
the next phase was learning about it s API and figure what s needed to
be modified next. Libvirt uses a XML file for setting Domain (in
libvirt s naming) definitions. So, Eucalyptus provides a Perl wrapper to
generate this file on runtime and allow the NC to invoke libvirt s API
to run the instance. Next task then, was adapt such script to support
ARM arch. The current script is taiolred for amd64 and i383. I worked on
that front and managed to get a script prototype, that can later be
improved and support more arch s.
Generating an adequate Libvirt s XML Domain definition file for ARM can
be a heroic task. There are many things to have in mind given the
diversity of the ARM processor and vendors. I focused on the versatile
flavour given that I was going to test the image I built on the first
place and it ran fine under qemu-system-arm.
The Perl script wrapper then was adapted for such configuration and it
can be tested independently by issuing the following command:
$ tools/gen_kvm_libvirt_xml arch arm
The arch parameter is what I implemented and it s intended to be
called from the kvm handler on instance creation. With the extensibility
in mind I ve created a hash to associate the arch with their
corresponding emulator.
our %arches = (
amd64 => /usr/bin/kvm ,
arm => /usr/bin/qemu-system-arm ,
);
Added the arch parameter to the GetOptions() function and used
conditions to tell whether the user is looking for a particular arch,
arm in our case. The most important parts to be considered are the
, and entries.
hvm
$local_kvm
There s also and that require to be tailored for arm.
root=0800
As output the script generates an XML template that can be adapted to
your needs and it could be used and tested with tools like Libvirt s
virsh. So that, it can be useful independently of Eucalyptus. I managed
to get to this point before the midterm evaluation.
Now that we had the XML wrapper almost done, the next task was to make
the handler call pass the arch as an argument to the script, so the
image is loaded with the proper settings and we are able to run it.
Eucalyptus doesn t have a arch field for NC s instances. So, after approaching
Eucalyptus developers, with whom we had already being interacting, I
settled to my proposal of extending the ncInstance struct with an arch
field (util/data.h). The ncInstance_t struct stores the metadata for the
instance being created. It s used for storing runtime data as well as
network configuration and more. It was indeed the right place to add a
new field. I did so by creating the archId field.
Now I needed to make sure the arch information is stored and later used on the
libvirt s call.
typedef struct ncInstance_t
char instanceId[CHAR_BUFFER_SIZE];
char imageId[CHAR_BUFFER_SIZE];
char imageURL[CHAR_BUFFER_SIZE];
char kernelId[CHAR_BUFFER_SIZE];
char kernelURL[CHAR_BUFFER_SIZE];
char ramdiskId[CHAR_BUFFER_SIZE];
char ramdiskURL[CHAR_BUFFER_SIZE];
char reservationId[CHAR_BUFFER_SIZE];
char userId[CHAR_BUFFER_SIZE];
char archId[CHAR_BUFFER_SIZE];
int retries;
/* state as reported to CC & CLC */
char stateName[CHAR_BUFFER_SIZE]; /* as string */
int stateCode; /* as int */
/* state as NC thinks of it */
instance_states state;
char keyName[CHAR_BUFFER_SIZE*4];
char privateDnsName[CHAR_BUFFER_SIZE];
char dnsName[CHAR_BUFFER_SIZE];
int launchTime; // timestamp of RunInstances request arrival
int bootTime; // timestamp of STAGING->BOOTING transition
int terminationTime; // timestamp of when resources are released (->TEARDOWN transition)
virtualMachine params;
netConfig ncnet;
pthread_t tcb;
/* passed into NC via runInstances for safekeeping */
char userData[CHAR_BUFFER_SIZE*10];
char launchIndex[CHAR_BUFFER_SIZE];
char groupNames[EUCA_MAX_GROUPS][CHAR_BUFFER_SIZE];
int groupNamesSize;
/* updated by NC upon Attach/DetachVolume */
ncVolume volumes[EUCA_MAX_VOLUMES];
int volumesSize;
ncInstance;
With that in place what was left was modifying the functions that
store/update the ncInstance data, and also the libvirt s function that
calls the gen_kvm_libvirt_xml script. I ve modified the following files:
- util/data.c
- node/handlers_kvm.c
- node/handlers. c,h
- node/test.c
- node/client-marshall-adb.c
Most important the allocate_instance() function that is in charge of setting up
the instance metadata a prepare it to pass it to the handler, then
hyphervisor trough Libvirt s API. The function now has a new archId
parameter as well, to keep coherence with the field name. It also
handles whether the field is set or not. We (Eucalyptus developers and
I) haven t settled wether to initialize this field with a default value
or not. I d stepped up and initialized with a NULL value for this string
type var.
if (archId != NULL)
strncpy(inst->archId, archId, CHAR_BUFFER_SIZE);
Stores the value only if it s passed to the function. This is essential to
don t break existing functionality and keep consistency for later releases.
With this we almost finished the part of extending Eucalyptus to support
arm images. It took quite long to get to that point. Next step: test.
As I mentioned in my previous report the current Eucalyptus packaging
has issues. From the beginning I ve approached the pkg-eucalyptus team,
who were very hepful. So I myself set me up and joined the pkg team with
commit access to the repo. Even the GSoC is not intended for packaging
labour, I needed to clear things on that side because what we want is
adoption so anyone, say a DSA or a user, could setup a cloud that has
support for running arm instances under amd64 arch.
Over the weeks between mid-term evaluation and past week I ve worked on
that front. Results were good, as you can see in the pkg-eucalyptus
mailing list and SVN repo.
Eucalyptus has a couple of issues that because of it s nature and complexity
triggered build errors. Those issues came from the Java side of the project and
few form the C-side. The C-part was a problem with AXIS2 Jar used to generate
stubs for later inclusion on build-time. There s no definitive solution to date
because one of the Makefile on gatherlog/Makefile issues a subshell call that
doesn t catches the AXIS_HOME environment variable. I ve worked around this
by defining it as a session env var from .zshrc. The other problem is related
to the Java libary versioning and most likely (we had to investigate further)
to groovy versioning.
As I explained Eucalyptus has 3 components written in Java. The Java code uses
extensively the AXIS2 library for implementing webservices calls, the groovy
language for JVM, and many other Java libaries. Eucalyptus open source version
ships both online and offline versions. The difference is basically that
the offline version ships the Java libaries JAR files inside the source code,
so you can build with all required deps.
Our former packaging used a file to list which Java libraries the package
depends on and it symlinked them using system wide installed packages. This
rather than helping out triggered a lot of problems I explaing with more detail
on the pkg-eucalyptus ml. I worked around this by telling the debian/rules
not to use such file to symlink and instead use the existing ones from
the source. The last weekend I managed to get to that point and built
sucessfully the software under Debian. I ve done that in order to have a
complete build-cycle for my changes, and then test. In fact today I ve
spotted and fixed a couple of bugs there.
What s available now
To date we have a Eucalyptus branch where my patches are sent. The level of
extension is nearly complete, more testing needed. One missing bit is to
change the UI so that the user selects which arch her image is and that
value is passed to the corresponding functions on the Node Controller
component.
We also have an ARM image that can be tested and later automated. I ve
recently learned about David Went s VMBuilder repo that extends Ubuntu s
vmbuilder to support image creationg. I might talk to him to adapt the tool
in order to support creation of arm images and use it to upload to an
Eucalyptus cloud.
I ve also created a wiki page with different bits on which I was working on.
I ve yet need to craft it to be more educational. I guess I m going to use
this report as a source :)
We also had more strong cooperation with the Eucalyptus project and
involved their developers on the packaging and this project, which is
something I consider a great outcome for this project and GSoC goals:
to attract more people to contribute to FLOSS. I ve also been contacted by
people from the Panda Project and we might cooperate also.
Future work
I plan to keep working on the project regardless the GSoC ends. Our goal
as Debian could be integrating this into the Eucalyptus upstream
branch. We have good relations with them so expect news from that side.
I also plan to keep working with the Eucalyptus team and related
software in Debian, given that I m familiarized with the tools and
projects.
I m also planning to advocate SoC for Debian in my faculty. To date we
had already 4 students, IIRC, that have participated in the past.
Project challenges and lessons
During the project I faced many challenges both in the personal side
and in the technical one. This part is a bit personal but I expect we
learn from this.
First challenge was the where do I sit problem. My project was
particular, indeed quite different from the others on it s nature. I had
to work on a software that is not Debian s and then come and say Hey we
have this nice tool for you, come test it! . So, indeed, my focus was on
the understanding of such tool, Eucalyptus, and not much on looking back
at Debian because I didn t feel I had something to advertise yet and
indeed I got most valuable/useful feedback from Eucalyptus developers
rather than Debian s, which for the case is OK IMHO. I had great
mentoring. This situation probably looked a bit of isolation from my
side, from Debian s POV, but it was not intended it s just how it needed
to be IMHO. I felt that I wasn t contributing directly to the
project. I ve spoke about such thing with Steffen a couple of times. I d
like to say that it was also my concern but I probably failed on
communicating this, second challenge.
Third challenge was more technical and related to the first one. Since I
was not writing any code for a Debian native package/program but instead for
other project I faced the situation of where to publish my
contributions. Eucalyptus had some issues on managing their FLOSS repo
and the current development one is outdated comparing to the
-src-offline 2.0.3 release. The project didn t fit pkg-eucalyptus repo
neither, even we thought to branch the existing trunk and ship my
contribs as quilt patches, but it involved a lot of more work and burn
on reworking something that has issues now. Later on time I branched the
bzr eucalyptus-devel Launchpad repo and synced there with the
src-offline code. I was slow to react on thing I can say.
Fourth challenge was also technical and still dealing with. I was looking to
setup a machine to set a cloud and test my branch. The Dean of my faculty
kindly offered a machine but I ve never managed to get things arranged
with the technical team. Steffen and I discussed about this and evaluated
options, at the end we settled on pushing adoption rather than showcasing it.
Since for adoption we need a use case, I ve spend some SoC money on hardware
for doing this. I should have news this week on that side :)
Finally, I d like to thank everyone on the Debian SoC team for this
opportunity to participate. I like to thank Steffen for his effort on
arranging things for me, Graziano from Eucalyptus for his advice on
working with the software code. The different people from Eucalyptus I
interacted with and showed interest on the project s sucess. My
classmates and teachers from the Computer Science School at UCSP in
Arequipa-Peru, friends from the computing community in Peru SPC and
finally my family. Thank you, I learned a lot, specially on the personal
side. As Randy Pausch put: The brick walls are there to give us a
chance to show how badly we want something.
Resources
Please see the Wiki page[1] for any reference to the project s resources.
1- http://wiki.debian.org/SummerOfCode2011/BuildWithEucalyptus/ProjectLog
Best regards,
Rudy
We've had sporadic Altus Metrum customer reports
about RF susceptibility issues with their
TeleMetrum installations. In almost
every case, these problems have been completely resolved by either making
sure the system battery has sufficient charge before launch, or through the
application of standard engineering techniques such as twisting wire pairs
to reduce differential coupling. However, even when every technique we could
think of had been applied, once in a while someone still had issues.
Around the time of LDRS this year, the incidence of such reports seemed to
increase. One customer, in particular, had an installation in which he
virtually always saw continuous resets of the board once his 54mm airframe
was put on a launch rail, and several customers reported seeing board resets
during ejection charge firing. Keith and I saw a board
reset during main charge firing happen in person at
NCR's Oktoberfest, and with a couple days available
to work together after that launch, we decided it was time to figure out
what was really going on.
Here's what we've learned.
In bench testing, it quickly became clear that the problem was the 3.3
volt power supply rail getting pulled down far enough to reset the CPU.
This most frequently happened during ejection
charge firing, when the input of the LDO regulator is pulled down by the
near-short presented by the e-match when a pyro FET is turned on. To keep
the 3.3 volt rail voltage up during firing, we include a 100uF bulk capacitor
on the regulator's output. In all of our prior bench testing, we never saw
the 3.3 volt rail droop significantly. Clearly something had changed... or
maybe several things had changed?
The first thing I wondered about was whether the new Kalman filtering code,
which requires more compute cycles from the processor, might be consuming
enough more power to pull the rail down faster during charge firing. After
poking around at it, though, we have no data to suggest the new code makes a
measurable difference in power consumption.
The next thing we pondered was that at least some of the e-matches we and
others are using in the hobby now come from the fireworks industry, where it
is apparently considered a feature for the match to retain continuity after
firing. This means the input of the LDO gets held down for longer than with
the e-matches we used to use and Quest Q2G2 igniters that open when fired.
But that still didn't make sense as the root cause, as we chose the FET
firing time such that even with a dead short across the igniter terminals,
the 3.3 volt rail wouldn't be pulled down far enough to cause trouble during
firing.
One of the big changes between v1.0 and v1.1 on TeleMetrum was that the newer
boards incorporate a better reset circuit. This helps ensure the GPS chip
always comes up running at power on, which was a problem at temperature
extremes with older boards. However, a side-effect of this change is that
a v1.1 board will reset any time the 3.3 volt rail drops below 3.15 volts,
whereas older boards didn't trip until a much lower voltage. So the recent
increase in reports might just be related to more v1.1 boards being placed
in service?
While experimenting on the bench, we observed that injecting RF energy into
the input of the LDO regulator had the effect of pulling down the output
voltage, presumably because the internal reference source accumulates charge
and is fooled into thinking the output is too high. Since our designs all
have the power switch contacts ahead of the LDO, the wires going out to the
switch and back are effectively an antenna... as are, to a lesser extent,
the wires going to the e-matches. There is some variability from part to
part in just how badly the LDO reacts. But by attaching a tuned length of
wire as an antenna to the LDO input and playing around, we were finally able
to reproduce the problem reliably on a test board at my bench!
On further analysis, we realized that the output of the USB battery charger
chip and the input of the LDO both expect a 1uF bypass cap to ground. At
some point, those looked redundant and we eliminated one of the
two. Unfortunately, we weren't internalizing the fact that the switch leads
were between the two caps, and the one we left was on the output of the charger
and not at the input of the LDO. Placing a suitable bypass cap right at the
input of the LDO turns out to have a truly dramatic effect on RF immunity!
Once we realized that RF getting into the LDO input was the problem, Keith
pointed out that we used to see "noise" in the accelerometer data on earlier
boards that was caused by the 3.3 volt rail moving slightly during radio
transmit, which we fixed with a hardware change on v1.1. We are now
convinced that this was at least partly related to RF coupling to the LDO
input, not just the change in power consumption on the LDO output. We
didn't realize what was going on in earlier testing because we often didn't
have ematches wired up, so RF coupling was minimal. But going back to
flights logged with v1.0 boards that included deployment, and studying the
magnitude of the "steps" in acceleration data observed when the transmitter
was on, Keith was able to compute the amount the 3.3 volt rail must have
sagged if the real acceleration wasn't changing... which in some cases was
as much as 180mV! We think this proves that RFI could cause the LDO to
drop its output voltage below the reset controller set point on v1.1 boards.
Based on these observations, I'm making two hardware changes for the next
version of TeleMetrum (version 1.2), and Keith is also making a software
change. We have tested all of these changes on real boards both on the
bench and in test flights, and the net effect is a major improvement in
the RF immunity of TeleMetrum.
The first hardware change is moving to a slightly lower trip voltage on the
reset controller. Instead of 3.15, the new part trips at 3.00 volts nominal.
This gives us more "headroom" to tolerate 3.3 volt rail droop during charge
firing, and will allow the board to operate longer on a given battery
charge. This change is not relevant for v1.0 and prior.
The second hardware change is adding an appropriate bypass capacitor right
at the LDO input. This requires a PCB update, but it's possible for me to
update existing production boards by adding an 0402 cap right across the
appropriate pins on the regulator chip.
The software change prevents our altimeters from turning on the radio
transmitter while an ejection charge is firing. Since the RF transmitter
draws substantial additional power, this should help keep the 3.3 volt rail
from drooping. This may not really matter, but it feels like the right
thing to do. This change will be part of our next stable firmware release.
We think most TeleMetrum customers need not worry about these updates. But
if you have seen odd resets on the rail or during ejection charge firing in
flight with a TeleMetrum v1.1, feel free to contact me about updating
your existing board to include these improvements.
Two days ago, I finally brought my system up-to-date with regard to the GNOME packages that I use. I was very reluctant because I am using a heavily customized gnome-panel and I waited until 
I recently purchased a Seagate FreeAgent GoFlex 1 TB hard drive. It is a 2.5" compact disk that you connect over USB. It draws its power from the USB controller. My previous Extern HDDs were all 2.5" laptop drives, for which I bought a USB enclosure. Those devices always worked perfect, as in:
Volker Theile is well known for his work on 
The first public release (Version 0.2, codenamed Ix) happened on 
I just 
Traditionally, when you start a program to view a document, for instance
A continuation of my 
I just download this tool 
This is just a quick post to remind myself in the morning, as soon as I've made it I intend to turn my computer off and leave it off until I can re-organize my office.
I've been using mutt for my email for the past few years. Nothing compares to the flexibility of procmail/sieve for organizing server-side mail, and then mutt is ideal for reading them.
With the addition of the mutt-patched sidebar mode you can even go for a few days before realizing you're not in a graphical environment. But one thing I do long for is the ability to execute scripts at various times.
Thus far I've not actually planned what I'd like to do, but as a starting point imagine being able to execute a hook when new mail arrives? Or when you send a message matching a pattern in some fashion?
There are some things out there, such as the various hacks which are designed to abort sending a message if you mention "See attachment" in a message body but fail to add one before sendign the message. These hacks generally abuse the sendmail configuration such that they're extremely ad-hoc and hard to chain/nest.
I've mellowed out over the years and I have no interest in attempting to write a mail-client (though at the same time how hard can it be? Just restrict yourself to using inotify on ~/Maildir and offload delivery to exim and you're almost done? I guess the hard part is the UI, though I do like the mutt + sidebar layout. Write the whole thing in some scripty language?)
I'll re-examine notmuch and gnus over the next week or two, but I suspect both will continue to disappoint in various ways.
Anyway, for the moment I'm just pondering. But threading is an obvious concern. Most current mutt hooks relate to the local folder, or the local message. If I were viewing a message in one directory and a new mail notification fired for a delivery to both ~/Maildir and ~/Maildir/.people.foo I'd need to either serialise them or thread them.
Ponder ponder.
In other news I've been doing more photography recently. Nothing cohesive except for my recent experiment with shooting a "street-girl" outdoors in falling light, but that was an interesting challenge and the results were sufficient to make me want to try shooting outdoors in an organized fashion again. (Some random images have been linked to from my wee 
As part of my development of a 
If you were trying to get rid of hiring discrimination (on grounds
irrelevant to the ability to do the job), you d think a good way to do it
would be to reduce the ability of the hiring manager to discriminate, by
restricting their access to irrelevant (but possibly prejudicial)
information. It s certainly what I might come up with as an early idea in a
brainstorming session.
I m not alone: