Am 24.03.2015 um 23:40 schrieb André Warnier:
Rainer Jung wrote:
Some numbers from a test here on RHEL 6, using Java 1.7.0_76 and TC
6.0.43, 7.0.59 and 8.0.20.
Measurement is taken directly after start (a) plus once after one
request to a non-existing page and two full GCs (b). Only manager was
deployed, not example webapps or docs. GC was run using "jcmd PID GC.run"
Numbers from "ps"
RSSa RSSb SZa SZb VSZa VSZb
tc6 62372 68336 272952 273532 1091808 1094128
tc7 63608 70456 271710 271978 1086840 1087912
tc8 72576 79140 272257 272525 1089028 1090100
Differences between TC6 and 7 marginal, differences between tc7 and 8
only noticable in RSS, around 9MB.
Numbers from jstat -gc. First "Capacity":
Semi Spaces start with 512KB and grow to 768.0KB (TC6), 896.0KB (tc7)
and 1024.0 (tc8). Those indicate increasing allocations, but are not
relevant for total memory use.
Edena Edenb Olda Oldb Perma Permb
tc6 4288.0 6656.0 10688.0 16320.0 21248.0 21248.0
tc7 4288.0 7168.0 10688.0 17904.0 21248.0 21248.0
tc8 6144.0 8640.0 15316.0 21316.0 21248.0 21248.0
Again this is capacity so including garbage and unused. We see that
Perm is unchanged. For all versions Eden grows by 2.4-2.9 MB due to
allocation activity. Numbers for tc6 and 7 are again very similar, tc8
numbers are slightly higher already after startup. Old (Tenured) grows
by about 6-7MB, again very similar for tc 6 and tc 7 and slightly
higher for TC 8.
Now for the used numbers after GC, which are more relevant (allocation
rates are another topic):
Edena Edenb Olda Oldb Perma Permb
tc6 2910.8 69.3 7231.7 7984.1 13923.0 14429.1
tc7 2326.1 73.8 8504.4 9661.2 13910.1 15340.1
tc8 203.7 60.9 10577.6 12599.7 16183.3 17653.8
So the live objects are Edenb+Oldb:
Edenb+Oldb
tc6 8053.4
tc7 9735.0
tc8 12660.6
And here we see some increase but the total amount of about 2MB
between tc 6 and 7 and about another 3 MB between 7 and 8 seems to be
not really problematic. The same holds true for perm, there's an
increase of about 1MB between 6 and 7 and 2 MB between 7 and 8.
Finally: where does the difference between RSS, Sz and the sum of heap
and eden come from?
Example for TC 8 case b:
RSS: 79140
SZ: 272525
Sum of RSS due to "smaps": 79088 so roughly consistent.
S0+S1+E+O+P capacity: 53252, but Rss 40872, so a delta of 38MB to RSS.
smaps entries that can be identified:
Type Size Rss
Perm 21248 17656
Old 21316 15040
Eden+S0+S1 10688 8176
libjvm.so 11732 7712 (read-only)
Then about 21 thread stack reservations, total Size 21676, total Rss
2804.
So the delta goes down to 38 - 7.7 - 2.8 = 28MB.
Some more segments, that I can't fully interprete are:
Size Rss from - to Perm File
8852 7648 7f2818000000-7f28188a5000 rw-p
8940 6496 7f2814000000-7f28148bb000 rw-p
51116 4064 7f2835270000-7f283845b000 rw-p
2496 2048 7f2835000000-7f2835270000 rwxp
1788 1732 7f283d443000-7f283d602000 r--s .../lib/rt.jar
3304 1444 7f282c000000-7f282c33a000 rw-p
784 784 7f283f1d9000-7f283f29d000 rw-p .../lib/amd64/server/libjvm.so
1564 648 3f6d200000-3f6d387000 r-xp /lib64/libc-2.12.so
536 524 7f2830000000-7f2830086000 rw-p
272 208 7f283f29d000-7f283f2e1000 rw-p
and those nearly make up the "missing" 28MB Rss (whatever they are).
...
But now, for the mere humans among us, what does it mean in terms of the
OP and his original question : why does Tomcat 7 seem to be using 70 MB
more memory at startup than Tomcat 6 ?
Is it :
- it doesn't matter. The numbers shown are wrong, and if you run 10
instances of Tomcat 7 at the same time, you will see that they are not
really using 700 MB more than before.
or
- it is normal and expected. Tomcat 7 - because of the new Servlet Spec
- needs to borogrove the watchamecalits, and this is using 70 MB more
heap than before. In return, you get a 25% performance improvement later..
or
- we have no clue. It does not happen on other machines, so there must
be something special on your machine, and to find out what we need heap
dumps.
or
- obviously some cleverer and definitive answer derived from Rainer's
exhaustive analysis abobe, and which is ? ................
From the above analysis, I get the impression that there is only really
a couple of MB additional memory used as one goes from Tomcat 6 to
Tomcat 7 and then to Tomcat 8.
Yes. Things might change if a real app comes into play. But again I
would not expect big differences if it is an app that only uses old
features, e.g. runs in old and new TC versions unchanged.
Note that I didn't do allocation measurements (bytes per second), only
memory demand. Allocation is a different story.
And that this can easily be explained by additional things/functionality
which each version does, compared to the previous one.
And it is so little memory, that IMHO it is not worth trying to break it
down into individual explanations.
But then, what could explain the 70 MB difference as shown by "top" ? Is
it really just illusory ?
Good question as always.
Currently the apples vs. pears risk is to high. I just wanted to find
out, whether there actually is a generally big difference or not. IMHO
there is not. I have the feeling there are to many unclear parts
concerning the OPs measurements here:
- were both tests done on the same system? Really the same? To get the
base numbers for differences one would have to ensure that.
- were both tests done on an otherwise completely idle system having no
memory pressure at all? Meaning also, that for both measurements only
the TC version under observation was running.
- were both tests done with a tomcat default configuration and
installation from the ASF downloads? Again that's necessary to get base
numbers.
- Assuming these conditions are met, can the OP produce numbers that can
be compared to mine (ps-numbers, jstat-numbers, doing two measurements,
one after startup, the other after one request plus two times Full GC)?
How do they differ?
From that point on one could make a plan to explain differences.
Regards,
Rainer
---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org
For additional commands, e-mail: users-h...@tomcat.apache.org
