No Huddle Offense

"Individual commitment to a group effort-that is what makes a team work, a company work, a society work, a civilization work."

Forget static callgraphs – Use Python & DTrace!

October 20th, 2011 • Comments Off on Forget static callgraphs – Use Python & DTrace!

Forget about static analyzed callgrahs! No more running the code closing it and then looking at the callgraph. With DTrace you can attach yourself to any (running) process on the (running/production) system and get life up to date information about what the programm is doing. No need to restart the application or anything. This works for most programming languages which have DTrace providers (like C, Java and Python :-)). All you need to know is the pid.

Based on the information you get from DTrace (using the Python consumer) you can draw life updating callgraphs of what is currently happening in the program. Not only is it possible to look at the callgraph but you can also look at the time it took to reach a certain piece of code to analyze bottle necks and the flow of the program:

$ pgrep python # get the pid of the process you want to trace
$ ./ 123456 # trace the program and create a callgraph

So if you would have the following Python code:

class A(object):

    def sayHello(self):
        return 'I am A'

class B(object):

    def __init__(self):
        self.a = A()

    def sayHello(self):
        return self.a.sayHello()

if __name__ == '__main__':
    print B().sayHello()

You would get the following life generated callgraph – the GUI can start, stop and restart tracing and get live updates as the DTrace probes fires:

Click to enlarge

The following screenshot was taken while looking into the printer manager:

Click to enlarge

DTrace for the win!

[Updated] Updated the screenshots.

Python traces Python using DTrace

October 19th, 2011 • Comments Off on Python traces Python using DTrace

Another example of how to use Python as a DTrace consumer. This little program traces a Python program while is runs and shows you the flow of the code. The output is displayed in a Treeview (An indent mean that python called another function – Stepping back means that the function returned) and when double clicking the source code is displayed (Would be nice to open pydev as well).

Click to enlarge

Another example of Python as a DTrace consumer: This small GUI gives an up to date view of the number of syscalls made by an executable. Since this GUI is a live up to date view you can watch the circles appear, grow and become smaller again 🙂

Click to enlarge

Now on to other things…maybe creating live animated callgraphs as your program runs? 😛

Python as a DTrace consumer – Part 2 walk the aggregate

October 8th, 2011 • Comments Off on Python as a DTrace consumer – Part 2 walk the aggregate

Yesterday I blogged about how to use Python as a DTrace consumer with the help of ctypes. The examples in there are very rudimentary and only captured the normal output of DTrace – not the aggregates.

The examples in the last post have been altered and now we let DTrace work for a few seconds and then walk the aggregate:

    # aggregate data for a few sec...
    i = 0
    chew = CHEW_FUNC(chew_func)
    chew_rec = CHEWREC_FUNC(chewrec_func)
    while i < 2:
        LIBRARY.dtrace_work(handle, None, chew, chew_rec, None)

        i += 1


    walk_func = WALK_FUNC(walk)
    # sorting instead of dtrace_aggregate_walk
    if LIBRARY.dtrace_aggregate_walk_valsorted(handle, walk_func, None) != 0:
        txt = LIBRARY.dtrace_errmsg(handle, LIBRARY.dtrace_errno(handle))
        raise Exception(c_char_p(txt).value)

The walk function is right now very simple but does work – please note the TODO 🙂

def walk(data, arg):
    Aggregate walker.
    # TODO: pickup the 16 and 272 from offset in dtrace_aggdesc struct...

    tmp = data.contents.dtada_data
    name = cast(tmp + 16, c_char_p).value
    instance = deref(tmp + 272, c_int).value

    print '+--> walking', name, instance

    return 0

When run the Python script will output (Would be fun to run this DTrace script with the help of Python – Python as a DTrace consumer tracing Python as DTrace provider :-P):

+--> In chew: cpu : 0
  +--> In out:  Hello World
+--> walking updatemanagernot 2
+--> walking mixer_applet2 4
+--> walking gnome-netstatus- 135
+--> walking firefox-bin 139
+--> walking gnome-terminal 299
+--> walking python2.7 545
Error 0

Overall this works pretty smoothly – but needs a lot of updating before it is production ready – Still it gives an rough overview that Python can be a simple DTrace consumer while using ctypes. So now Python can be consumer and provider for DTrace *happy days* 🙂

The code (examples) have been updated on github.

Python as a DTrace consumer using libdtrace

October 7th, 2011 • 3 Comments

So DTrace is awesome – You can create nice analytics with it like Joyent does. But what I wanted is to access the output/aggregations from DTrace using Python to be able to parse the output as it comes. Using DTrace to Monitor Python or zones is already easy 🙂 Doing so it might be able to express up to date information to a (Cloud)Client (using of course sth. like OCCI)

So all I need is a Python based DTrace consumer which uses the dtrace library. To start with let’s reading the comments in the file /usr/include/dtrace.h – most notable is:

Note: The contents of this file are private to the implementation of the Solaris system and DTrace subsystem and are subject to change at any time without notice.

So we are already on our own – Next to that there is very limited documentation on writing DTrace consumers. A few quick searches might help you find some information. Probably the most up to date is to look into Bryan Cantrill’s consumer for node.js: And as mentioned a Python based consumer for libdtrace should be the goal of all this – so peeking to how others do it is probably a good idea :-P. For now let we will focus on a simple Hello World example.

First we need to understand how to use libdtrace – So let’s take a look at this diagram:


Following this life-cycle we can easily create some C code which will interface nicely with libdtrace. But since we can use it in C we can also use Python and ctypes to access the library. Here it is where we start the fun part.

To start with we will try to execute the following D script. It does nothing more than printing Hello World when executed using the dtrace command. But the output of this trace should now be received in Python code – so the output could be evaluated later on:

dtrace:::BEGIN {trace("Hello World");}

Now let’s write some python code – first we need to wrap some Structures which are defined in the dtrace.h file. Namely we will need dtrace_bufdata, dtrace_probedesc, dtrace_probedata and dtrace_recdesc. Since this is a blog post please refer to the source code at github for more details. We also need to define some types for callback functions – since we need a buffered writer, chew and chewrec functions as shown in the previous diagram:


def chew_func(data, arg):
    Callback for chew.
    print 'cpu :', c_int(data.contents.dtpda_cpu).value
    return 0

def chewrec_func(data, rec, arg):
    Callback for record chewing.
    if rec == None:
        return 1
    return 0

def buffered(bufdata, arg):
    In case dtrace_work is given None as filename - this one is called.
    print c_char_p(bufdata.contents.dtbda_buffered).value.strip()
    return 0

The function called buffered will eventually write the Hello World string later on – The chew function will print out the CPU id.

With the basic stuff available new can load the libdtrace library and start doing some magic with it:



Now all there is left to do is follow the steps described in the diagram. First step is to get an handle an set some options:

    # get dtrace handle
    handle = LIBRARY.dtrace_open(3, 0, byref(c_int(0)))

    # options
    if LIBRARY.dtrace_setopt(handle, "bufsize", "4m") != 0:
        txt = LIBRARY.dtrace_errmsg(handle, LIBRARY.dtrace_errno(handle))
        raise Exception(c_char_p(txt).value)

Setting the bufsize option is important – otherwise DTrace will report an error. Now we’ll register the buffered function which we wrote in Python and for which we have a ctypes type:

    buf_func = BUFFERED_FUNC(buffered)
    LIBRARY.dtrace_handle_buffered(handle, buf_func, None)

Now we will compile the D script and run it:

    prg = LIBRARY.dtrace_program_strcompile(handle, SCRIPT, 3, 4, 0, None)

    # run
    LIBRARY.dtrace_program_exec(handle, prg, None)

If this exists correctly (The C file in the github repository has all the checks for the return codes in it) we can try to get the Hello World. The chew and chewrec functions are also implemented in Python and can now be registered.

If the second argument on the dtrace_work function is None DTrace will automatically use the buffered callback function described two steps ago. Otherwise a filename needs to be provided – but we wanted to get the Hello World into our Python code:

    # do work
    chew = CHEW_FUNC(chew_func)
    chew_rec = CHEWREC_FUNC(chewrec_func)
    LIBRARY.dtrace_work(handle, None, chew, chew_rec, None)

And last but not least we will print out any errors close the handle on DTrace:

    # Get errors if any...
    txt = LIBRARY.dtrace_errmsg(handle, LIBRARY.dtrace_errno(handle))
    print c_char_p(txt).value

    # Last: close handle!

Now this all isn’t perfect and not ready at all (especially the naming of functions could be updated, a nice abstraction layer be added, etc) – but it should give a nice overview of how to write DTrace consumers. And for the simple example here both the C and Python code at the previously mentioned github repository do seem to work – and do in fact output:

cpu : 0
Hello World
Error 0

So maybe it’s time to combine pyssf (an OCCI implementation), pyzone (Manage zones using Python) and python-dtrace for monitoring and create a nice ‘dependable’ (Not my idea – these are the words of Andy) something

OpenIndiana zones & DTrace

July 6th, 2011 • Comments Off on OpenIndiana zones & DTrace

Let’s assume we want to create a Solaris zone called foo on an OpenIndiana box. This post will walk you to all the steps necessary to bootstrap and configure the zone, so it’s ready to use without any user interaction. Also briefly discussed is how to limit the resources a zone can consume.

7 Steps are included in this mini tutorial:

  1. Step 1 – Create the zpool for your zones
  2. Step 2 – Configure the zone
  3. Step 3 – Sign into the zone
  4. Step 4 – Delete and unconfigure the zone
  5. Step 5 – Limit memory
  6. Step 6 – Use the fair-share scheduler
  7. Step 7 – Some DTrace fun

Step 1 – Create the zpool for your zones

First a pool is created and mounted to /zones. Deduplication is activated for this pool & a quota is set – so the zone has a space limit of 10Gb.

zfs create -o compression=on rpool/export/zones
zfs set mountpoint=/zones rpool/export/zones
zfs set dedup=on rpool/export/zones

mkdir /zones/foo
chmod 700 /zones/foo

zfs set quota=10g rpool/export/zones/foo

Step 2 – Configure the zone

A zone will be configured the way that it has the IP (Nope – DHCP doesn’t work here :-)) and uses the physical device rum2. Otherwise the configuration is pretty straight forward. (TODO: use crossbow)

zonecfg -z foo "create; set zonepath=/zones/foo; set autoboot=true; \
 add net; set address=; set defrouter=; set physical=rum2; end; \
 verify; commit"
zoneadm -z foo verify
zoneadm -z foo install

Too ensure that when we boot the zone everything is ready to use without any additional setups, a file called sysidcfg is placed in the /etc of the zone. This will make sure that when we boot all necessary parameters like a root password, the network or the keyboard layout are automatically configured. Also the host’s resolv.conf is copied to the zone (this might not be necessary if you have a properly setup DNS server – than you can configure that DNS server in the sysidcfg file – mine does not know the hostname foo so that is why I do it this way) and the nsswitch.conf file is copied so it’ll use the resolv.conf file. Finally the zone is started…

echo "
network_interface=PRIMARY {hostname=foo
" &> /zones/foo/root/etc/sysidcfg

cp /etc/resolv.conf /zones/foo/root/etc/
cp /zones/foo/root/etc/nsswitch.dns /zones/foo/root/etc/nsswitch.files

zoneadm -z foo boot

To create a password you can use the power of Python – the old way of copying the passwords from /etc/shadow doesn’t work on newer Solaris boxes since the value of CRYPT_DEFAULT is set to 5 in the file /etc/security/crypt.conf:

python -c "import crypt; print crypt.crypt('password', 'aa')"

Step 3 – Sign into the zone

Now zlogin or ssh can be used to access the zone – Note that the commands mpstat and prtconf will show that the zone has the same hardware configuration as the host box (zfs list – will show that disk space is already limited). In the next steps we want to limit those…

Step 4 – Delete and unconfigure the zone

First we will delete the zone foo again:

zoneadm -z foo halt
zoneadm -z foo uninstall
zonecfg -z foo delete

Step 5 – Limit memory

Following the steps above just change the configuration of the zone and add the capped-memory option. In this example it’ll limit the memory available to the zone. When running prtconf it’ll show that the zone only has 512Mb RAM – mpstat will still show all CPUs of your host box.

zonecfg -z foo "create; set zonepath=/zones/foo; set autoboot=true; \
 add net; set address=; set defrouter=; set physical=rum2; end; \
 add capped-memory; set physical=512m; set swap=512m; end; \
 verify; commit"

Step 6 – Using Resource Pools

While using resources pools it is possible to create a resource pool for a zone which only has one CPU assigned. Use the pooladm command to configure a pool called pool1:

poolcfg -c 'create pset pool1_set (uint pset.min=1 ; uint pset.max=1)'
poolcfg -c 'create pool pool1'
poolcfg -c 'associate pool pool1 (pset pool1_set)'

pooladm -c # writes to /etc/pooladm.conf

To restore the old pool configuration run ‘pooladm -x‘ and ‘pooladm -s

Now just configure the zone to use and associate it with the pool:

zonecfg -z foo "create; set zonepath=/zones/foo; set autoboot=true; \
 set pool=pool1; \
 add net; set address=; set defrouter=; set physical=rum2; end; \
 add capped-memory; set physical=512m; set swap=512m; end; \
 verify; commit"

Running mpstat and prtconf in the zone will show only one CPU and 512Mb RAM.

Step 6 – Use the fair-share scheduler

Also if you have several zones running in one pool you want to modify the pool to use FSS – so a more important zone gets privileged shares:

poolcfg -c 'modify pool pool_default (string pool.scheduler="FSS")'
pooladm -c
priocntl -s -c FSS -i class TS
priocntl -s -c FSS -i pid 1

And during the zone configuration define the rctl option – This example will give the zone 2 shares:

zonecfg -z foo "create; set zonepath=/zones/foo; set autoboot=true; \
 add net; set address=; set defrouter=; set physical=rum2; end; \
 add capped-memory; set physical=512m; set swap=512m; end; \
 add rctl; set name=zone.cpu-shares; add value (priv=privileged,limit=2,action=none); end; \
 verify; commit"

Step 7 – Some DTrace fun

DTrace can ‘look’ into the zones – For example to let DTrace look at the files which are opened by process within the zone foo you can simply add the predicate ‘zonename == “foo”‘:

pfexec dtrace -n 'syscall::open*:entry / zonename == "foo" / \
  { printf("%s %s",execname,copyinstr(arg0)); }'

I was researching this stuff to create a Python module to configure and bootstrap zones so I can monitor the zones & their previously created SLAs.