Hi,
This patchset tries to optimize ebpf hash map, and follows
the ideas:
1) Both htab_map_update_elem() and htab_map_delete_elem()
can be called from eBPF program, and they may be in kernel
hot path, so it isn't efficient to use a per-hashtable lock
in this two helpers, so this patch converts the lock into
per-bucket bit spinlock.
2) kmalloc() is called in htab_map_update_elem() for allocating
element, together with one global counter for tracking how many
elementes have been allocated. kmalloc is often a bit slow,
and the global counter doesn't scale well. This patch pre-allocates
one element pool and uses percpu ida for runtime element allocation/free,
and the global counter is removed too with this approach.
With this patchset, looks the performance penalty from eBPF
decreased a lot, see the following test:
1) run 'tools/biolatency' of bcc before running block test;
2) run fio to test block throught over /dev/nullb0,
(randread, 16jobs, libaio, 4k bs) and the test box
is one 24cores(dual sockets) VM server:
- without patchset: 607K IOPS
- with this patchset: 1332K IOPS
- without running eBPF prog: 1492K IOPS
include/linux/rculist.h | 55 +++++++++++
kernel/bpf/hashtab.c | 247 ++++++++++++++++++++++++++++++++++++++----------
2 files changed, 252 insertions(+), 50 deletions(-)
Thanks,
Ming
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
