Peter E. Abresch Jr. - at Pepco wrote:
I have analyzed the LINUXP01 memory usage over last week. I noticed that
between the hours of 04:00 – 04:30 AM every morning is when the largest
swap increase occurs. 04:15 AM is when some Linux housekeeping starts.
These tasks complete in about 3-5 minutes. These tasks require memory
and Linux will swap out memory not being used. The memory that is being
swapped out is from WebSphere. This is normal Linux process. What is
interesting is that this memory is never swapped back in. As the first
graph depicts, we see a jump in swap space usage every morning during
the housekeeping process. Every day shows another increase in swap space
utilization and a corresponding decrease in WebSphere resident memory
size about the same time as shown in the second graph.
The swap space usage cannot be used as indication whether or not pages
are swapped back in: when a page gets swapped out for the first time,
it gets a swap slot assigned. This is reflected as swap usage
increase. Once it gets swapped back in, it is treated as swap cache:
we have the page in memory _and_ on disk. This has two advantages:
- if the page does not get changed ("dirty"), we don't need to swap it
out again - we can just discard it because we have an up-to-date copy
on disk.
- if we need to swap it out again, the page does already have a
location on disk assigned which removes overhead for housekeeping swap
pages when swapping the same page in and out over and over again.
If you look at /proc/meminfo the value "SwapCached:" indicates how
much was swapped out before and was brought back in. If you substract
SwapCached from the swap usage, you get the number you're looking for:
how much does reside on disk that is not present in memory.
Since the major portion of the swap space utilization is never brought
back in, it is obviously not needed. Is this indicative of too much Java
heap size which never forces garbage collection? Is there a “memory
creep” condition in WebSphere, WebSphere’s JAVA, or one of the
applications running under WebSphere? Are the trees blocking my view of
the forrest?
>
> What can I conclude from this? All comments and suggestions are
welcome.
> The Linux guest size is 768 meg and the Java heap size is min 50
and 700
> max.
I don't know about the java internal housekeeping, therefore I cannot
do a well-founded statement on this question. I would expect that the
jave runtime does page-in all of the memory over time when workload
demands it.
so long,
Carsten
--
Carsten Otte has stopped smoking: Ich habe in 10 Monate, 2 Tage und 20
Stunden schon 1.472,80 Euro gespart anstatt 6.136,70 Zigaretten zu kaufen.
----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390