[Oisf-devel] aligning data structures ...

[chetan loke]

Hello,

Some low hanging fruits if not done already -

I think it would be better if we align the _Packet struct on a
cache-line boundary.
This will avoid cache pollution amongst multiple threads.

So how do we do it in a platform(32/64 bit , different CPU models etc)
agnostic way?

1) Existing definition

typedef struct _Packet {
...
} Packet;


2) New proposed definition:

#define suricata_pkt_alignment __attribute__((__aligned__(_CLS)))


typedef struct _Packet {

} Packet suricata_pkt_alignment;

Also good to align counter etc structs.

Now the million dollar question - how do we get _CLS?

In the makefile add the following:

SURICATA_DEFINE_FLAGS=

...

PLATFORM_CLS=$(shell getconf LEVEL1_DCACHE_LINESIZE)

SURICATA_DEFINE_FLAGS := -D_CLS=$(PLATFORM_CLS) $(SURICATA_DEFINE_FLAGS)

example - CFLAGS=-Wall -Werror -g -ggdb -c $(SURICATA_DEFINE_FLAGS)


3) I would also recommend using posix_memalign as opposed to a 'malloc'
    while creating the packetpool during init-time in suricata.c:

posix_memalign((void
**)(&queue_ptr),sysconf(_SC_PAGESIZE),(SIZE_OF_PACKET)*(max_pending_packets))

malloc all the packets once and avoid the for loop.

4) Checking a global flag(even if page/cache aligned etc) to see if we
need to stop the suricata engine kill's performance.
One trick to get around this global flag is for the main thread to
shutdown the interface[s] which is[are] receiving packets.


If someone can please dump a short/quick(no fancy diagrams needed)
call-flow on the website(or in the doc section) then it will be
helpful for new comers to start contributing in less than a day, test
all these low hanging fruits and submit patches.


Regards
Chetan Loke