- 22 Jul, 2014 14 commits
-
-
Morris Jette authored
Conflicts: NEWS
-
Morris Jette authored
-
David Bigagli authored
HealthCheckProgram for nodes in any other states than IDLE.#978
-
Levi Morrison authored
-
Morris Jette authored
Conflicts: NEWS
-
Morris Jette authored
-
Morris Jette authored
Unload job tables rather than windows at job end. The table unload also unloads job tables.
-
David Linden authored
The code in question is allocating a buffer to hold a header and a data payload then forwarding it on a socket. The first thinko is that it sizes the buffer based on the data payload, which won't be enough for the header+data. For the second thinko, when it realizes there isn't enough room, it adds the data size to the _capacity_ of the buffer, effectively allocating twice as much memory as needed. (Aesthetically, I dislike adjusting the 'size' field of the structure to declare the buffer is larger before the buffer is actually made larger, thus the 'new_size' variable.) This patch corrects those thinkos. Motivation: As we all discovered with host parsing, xmalloc() of large data can be time-consuming, so avoiding it is preferred. And in this case avoiding it for 2x the memory needed. This is also a place where the xrealloc_nz() could be used, since the memory is immediately written into. As you are probably aware, as peer counts creep into the thousands, this data payload creeps into the megabytes, so this can be a noticeable latency. An alternate implementation is to guess the max size of the header, and allocate header+data and hope the realloc isn't needed. I don't do that for a couple reasons. If the guess is wrong, the worst case is that realloc will end up copying the buffer (full size, yet not without the data payload in it) to a new larger buffer, and I really dislike that prospect. Better is to assemble the small header, and let realloc copy the small buffer into the huge new space.
-
Morris Jette authored
Modify the src/common/pack.c functions to use the new memory management functions and not clear the memory allocated.
-
Morris Jette authored
Added new internal Slurm functions xmalloc_nz() and xrealloc_nz(), which do not initialize the allocated memory to zero for improved performance.
-
Danny Auble authored
super heavy systems to be able to only commit every once in a while instead of after every transaction. This is much faster, but can introduce a very small window in which if something happens that makes the slurmdbd end irregularly some data loss is possible.
-
Danny Auble authored
it is much faster than doing a commit every time.
-
Morris Jette authored
-
Morris Jette authored
switch/nrt - Do not explicitly unload windows for a job on termination, only unload its table (which automatically unloads its windows).
-
- 21 Jul, 2014 13 commits
-
-
Danny Auble authored
-
Nathan Yee authored
-
Nathan Yee authored
-
Danny Auble authored
-
Levi Morrison authored
-
Danny Auble authored
error... error: '0' flag ignored with '-' flag in gnu_printf format [-Werror=format=] printf("%0-*llX ", abs_len, (long long unsigned) value); There is probably be better way to fix it, I just wasn't sure how.
-
Danny Auble authored
-
Danny Auble authored
-
Danny Auble authored
-
Danny Auble authored
-
Danny Auble authored
-
Danny Auble authored
-
Levi Morrison authored
-
- 19 Jul, 2014 3 commits
-
-
Morris Jette authored
-
Morris Jette authored
-
David Bigagli authored
-
- 18 Jul, 2014 10 commits
-
-
Danny Auble authored
-
Danny Auble authored
-
Ryan Cox authored
-
Danny Auble authored
-
Danny Auble authored
-
Ryan Cox authored
levels of accounts with fairshare=parent
-
Danny Auble authored
-
Ryan Cox authored
fairshare=parent
-
Danny Auble authored
in the association tree with fairshare=parent
-
Danny Auble authored
-