Thanks for the feedback.
Just so you know you aren't crazy, I tested it on my EEE PC and I am
seeing similar results as you. It is taking around 18 seconds with an
empty cache and about 23 seconds with a populated cache.
In the end, though, the patch does show improvement over the non-patched
version, so it is probably still worthwhile...but very counter intuitive.
-> richard
On 10/5/10 12:21, Pierre De Rop wrote:
Hello Richard,
I progress on my testing, and I understand now why I did not get the same
results than you ...
I did two tests: one test on a machine with a relatively slow disk, and
another test on a
host whose disk is faster.
I did the test with the 213 bundles from glassfish 3.0.1
(glassfishv3/glassfish/modules/*.jar)
Test on the host with the slow disk:
========================
I first measured the time needed to open a single file when disk buffers are
flushed: 32 millis.
(I did this to check that the disk io speed).
Here are the results of the tests:
* with felix 3.0.3 ->
time to load an empty felix cache (with disk buffers flushed) = 9.2 millis
time to load an existing felix cache (with dsk buffers flushed) = 14.9
millis
* with felix trunk (optimized version that you uncommitted yesterday) ->
time to load an empty felix cache (with disk buffers flushed) = 9.1 millis
time to load an existing felix cache (with dsk buffers flushed) = 12.3
millis
(here, we notice that it takes a longer time to load an existing felix
cache, than an empty one, either with felix 3.0.3 or optimized felix trunk).
Test on the host with the faster disk:
==========================
Time needed to open a single file when disk buffers are flushed: 1 millis.
* with felix 3.0.3 ->
time to load an empty felix cache (with disk buffers flushed) = 8.1 millis
time to load an existing felix cache (with dsk buffers flushed) = 5.3 millis
* with felix trunk (optimized version that you recently uncommited) ->
time to load an empty felix cache (with disk buffers flushed) = 7.9 millis
time to load an existing felix cache (with dsk buffers flushed) = 4.7 millis
(here, loading an existing felix cache is faster then an empty felix cache,
as expected).
So, in any cases, the optimized version of the trunk is faster than the
2.0.3.
However, with a low disk IO rate, then it seems that it takes a longer time
to load an existing felix cache
than an empty one.
/pierre
On Tue, Oct 5, 2010 at 3:13 PM, Richard S. Hall<[email protected]>wrote:
On 10/5/10 2:59, Felix Meschberger wrote:
Hi,
On 04.10.2010 22:26, Richard S. Hall wrote:
On 10/4/10 12:31, Pierre De Rop wrote:
This is surprising, because I double checked (I did the same kind of
tests
of differents machines), and I always see that loading a populated
cache is
a bit longer than an empty cache.
Ok, it's confusing for now, I will do further investigation and will get
back to you later, if I find something.
Let's keep investigating this.
In the meantime, I decided to rollback the change for the next release
so we can have more time to investigate it because right now I need to
get a bug fix out. So, we won't introduce any bundle cache structure
changes in 3.0.4, but maybe in 3.0.5.
This sounds like a minor-release-number-worthy change ;-)
I'm not sure that's necessary. It is still backward compatible, but it
isn't forward compatible.
-> richard
Regards
Felix
-> richard
On Mon, Oct 4, 2010 at 5:05 PM, Richard S.
Hall<[email protected]>wrote:
On 10/4/10 10:52, Richard S. Hall wrote:
Thanks for the installer bundle, I did some tests locally on my Mac
using
230 bundles from GlassFish, this is what I see (note: the "purge"
command
flushes disk buffers):
Empty cache
-----------------
[heavyweight main]$ rm -rf felix-cache; purge; date; java -jar
bin/felix.jar
Mon Oct 4 10:43:09 EDT 2010
2010-10-04 10:43:24,541 - BundleInstaller starting and waiting for
FrameworkEvent.STARTED event ...
2010-10-04 10:43:24,822 - Framework started: Installing bundles from
./load dir
2010-10-04 10:43:54,464 - Done.
Populated cache
-----------------------
[heavyweight main]$ purge; date; java -jar bin/felix.jar
Mon Oct 4 10:44:49 EDT 2010
2010-10-04 10:45:12,212 - BundleInstaller starting and waiting for
FrameworkEvent.STARTED event ...
2010-10-04 10:45:13,340 - Framework started: Installing bundles from
./load dir
2010-10-04 10:45:13,342 - Done.
You can see that the empty cache scenario takes about 45 seconds,
whereas
the populated cache takes about 24 seconds. So, on my machine, it
takes a
lot less time to re-load cached bundles...could be a function of my
disk
speed, I suppose, but it is what I would expect.
Not sure how to proceed, other than trying to get more samples.
Interestingly, I tried the same test using the 3.0.3 release (the
above was
with trunk) and I got the follow results:
Empty cache (3.0.3)
-----------------
rm -rf felix-cache; purge; date; java -jar bin/felix.jar
Mon Oct 4 10:58:23 EDT 2010
2010-10-04 10:58:43,340 - BundleInstaller starting and waiting for
FrameworkEvent.STARTED event ...
2010-10-04 10:58:43,557 - Framework started: Installing bundles from
./load
dir
2010-10-04 10:59:03,373 - Done.
Populated cache (3.0.3)
-----------------------
purge; date; java -jar bin/felix.jar
Mon Oct 4 10:59:50 EDT 2010
2010-10-04 11:00:24,738 - BundleInstaller starting and waiting for
FrameworkEvent.STARTED event ...
2010-10-04 11:00:26,736 - Framework started: Installing bundles from
./load
dir
2010-10-04 11:00:26,738 - Done.
So, the populated cache takes nearly as long as the empty cache (36
seconds
compared to 40 seconds)...this also makes sense, since the trunk has
patches
specifically designed to improve re-loading cached bundles.
-> richard
On Fri, Oct 1, 2010 at 11:30 PM,<[email protected]> wrote:
On Fri, 1 Oct 2010 23:25:35 +0200, Pierre De Rop<
[email protected]>
wrote:
On Fri, Oct 1, 2010 at 10:57 PM, Richard S. Hall
<[email protected]>wrote:
One thing, was your performance comparison based on using
the new
cache
or
old?
I assume you are using a newly created cache for the framework
snapshot
and
[obviously] an old cache for the 3.0.2, correct?
yes:
old cache with 3.0.2, and new cache with trunk.
but I also checked that new fwk/trunk is properly working when
being
started
with the old cache (from 3.0.2)
You should also be testing an already created cache (i.e., a
framework
restart), not an empty cache.
In the previous test, I compared with already created cache,
not
with
empty
caches.
With empty caches, here are the results:
- old fwk 3.0.2 / fresh empty cache -> 15.8 sec. (this is
surprising:
I
would expect to get a longer delay, since the cache is
empty when
I
start
the fwk)
2010-10-01 23:10:17,403 Starting Felix 3.0.2
2010-10-01 23:10:23,885 Starting cluster node "test"
2010-10-01 23:10:33,208 Cluster node "test" ready
- new fwk / fresh empty cache ->15.2 sec
2010-10-01 23:15:46,2 Starting Felix 3.1.0.SNAPSHOT
2010-10-01 23:15:52,417 Starting cluster node "test"
2010-10-01 23:16:01,225 Cluster node "test" ready
hmmm, I don't understand why the fwk startup time is better
when the
cache
is empty (15.2 sec) than when the cache is non-empty (21.2 sec
from
previous
mail)
Ok, I will do more tests ... may be I did something wrong ? ...
to be
continued ...
Hmm. Yeah, that seems a little odd. In that case, you should
just
delete
your cache each time to get better startup performance! ;-)
Definitely let me know what you find out. I too would expect
framework
restarts to be faster since we don't have to copy the JAR files
first.
-> richard
/pierre
Is that what you are doing? Just want to make sure I
understand what
is
being measured. :-)
-> richard
On 10/1/10 16:43, Pierre De Rop wrote:
Hi Richard,
I tried our app server with the trunk (old cache with the new
fwk/trunk)
->
everything works ok
(I install 195 bundles, and start 41 bundles)
With new cache/fwk trunk -> it's also ok.
I also compared the performance of the startup time between
felix
2.0.2
and
felix trunk.
To do this, I managed to ensure that my linux IO buffers
cache are
empty
before doing each test, by doing this:
sync
echo 1> /proc/sys/vm/drop_caches
echo 2> /proc/sys/vm/drop_caches
echo 3> /proc/sys/vm/drop_caches
This ensures that my linux has released all buffer cache
disk, like
it
is
the case when I boot my host (cold start).
Now, for felix 3.0.2, I load my server in around *21,1 *seconds
(it's
slow
because all disk buffers are empty):
2010-10-01 22:15:07,880 Starting Felix 3.0.2
2010-10-01 22:15:18,121 Starting cluster node "test"
2010-10-01 22:15:28,966 Cluster node "test" ready
and with felix/trunk, indeed, the server starts a little bit
quicker,
but
not so much: around *in 20,3* seconds (I also flushed disk
buffers
before
starting the test):
2010-10-01 22:17:47,505 Starting Felix 3.1.0.SNAPSHOT
2010-10-01 22:17:56,815 Starting cluster node "test"
2010-10-01 22:18:07,798 Cluster node "test" ready
/pierre
On Fri, Oct 1, 2010 at 5:39 PM, Richard S. Hall<
[email protected]
wrote:
p.s. I deploy snapshots of the everything, including the
convenience
framework distribution or you can build from trunk.
On 10/1/10 11:37, Richard S. Hall wrote:
Hey everyone,
I've made some changes to the bundle cache layout and
handling to
improve
performance and clean up code. I've tried to maintain
backward
compatibility
with old caches, but I'd appreciate is someone could try it
out
on
their
caches to see if it works.
** Back up your caches just in case! **
Note that the new cache format will not work with older
frameworks.
So
you
cannot use a cache created with this new framework with previous
frameworks;
however, previously created caches should mostly continue
to work
between
the two although there will be a loss of fidelity since
changes
to
state
are
only saved to the new way or the old way depending on if
you are
running
on
the new framework or the old one.
For the curious, I've combined five bundle cache files (id,
location,
state, start level, last modified, and refresh count) into a
single
file
and
try to avoid excessive file accesses. This appears to improve
startup
time
when you have a cache with lots of bundles, but your mileage
will
vary
based
on platform and other factors. Although, you won't necessarily
see
any
improvements if you are using an old cache, since it will
revert to
the
old
way.
Thanks.
-> richard
