[lustre-discuss] Lustre caching, chapter 2
John Bauer
bauerj at iodoctors.com
Thu Mar 26 08:29:28 PDT 2026
All,
An update on the previous lustre-discuss post. The closing of the first
file after writing, dumping from cache, and reopening with O_RDONLY,
does make it immune to the stat()s from another client node.
Now to the real question I have concerning the Lustre caching for the
reproducer program. I have changed the reproducer to :
1: write the first file, sim.dat.A.
2: force a memory dump of the OSCs ofA by doing O_DIRECT reads of the
first stripe block on each OSC. ( a trick I recently learned )
3: close the fd for A and reopen with only O_RDONLY.
4: read A.
5: write the larger second file, sim.dat.B
6: using the same fd, without closing B, read B forwards and backwards.
7: close both files.
This can all be see in this first image.
Why do the OSCs forA retain their memory even though the data is
used/touched only once, while the OSCs for B are fighting for caching
memory and their data is used/touched 3 times?
https://www.dropbox.com/scl/fi/flpvi9dnvftbtpjquleef/3461_fpa.png?rlkey=ux7zko8igh6loosrfs05ryb1c&st=mtpixhyy&dl=0
Another plot of interest relating to this run is the amount of memory
for filePages on each node(socket) of the system. This happens to be an
8 node system. When A is read back in it is placed on nodes 1,2, and 4.
While B is later fighting for caching memory, it also has filePages on
node 4. How can A hang on to the memory when its pages's LRU is
certainly less than or equal to those for B. The memory behavior can't
be a node thing either as both files are on node 4 for much of the run,
but A never gets pushed out of node 4. I should also point out that the
aggregate of the filePages on the 8 nodes maps directly to the cached
memory from meminfo. Not sure why it was decided to call the node thing
filePages and the meminfo thing cached in /proc.
https://www.dropbox.com/scl/fi/gcpbct21e0knx5d420njo/3461_nodemem.png?rlkey=c47nl06a5hoqswh62ivx9uxdh&st=za8ec97b&dl=0
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20260326/b0bc2328/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 3461_fpa.png
Type: image/png
Size: 69821 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20260326/b0bc2328/attachment-0002.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 3461_nodemem.png
Type: image/png
Size: 37903 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20260326/b0bc2328/attachment-0003.png>
More information about the lustre-discuss
mailing list