Re: Re: [PATCH] ksm : use checksum and memcmp for rb_tree

From: Kyeongdon Kim
Date: Mon Nov 06 2017 - 20:39:30 EST


Sorry, re-send this email because of the Delivery failed message (to linux-kernel)

On 2017-10-30 ìí 10:22, Timofey Titovets wrote:
2017-10-30 15:03 GMT+03:00 Kyeongdon Kim <kyeongdon.kim@xxxxxxx>:
> The current ksm is using memcmp to insert and search 'rb_tree'.
> It does cause very expensive computation cost.
> In order to reduce the time of this operation,
> we have added a checksum to traverse before memcmp operation.
>
> Nearly all 'rb_node' in stable_tree_insert() function
> can be inserted as a checksum, most of it is possible
> in unstable_tree_search_insert() function.
> In stable_tree_search() function, the checksum may be an additional.
> But, checksum check duration is extremely small.
> Considering the time of the whole cmp_and_merge_page() function,
> it requires very little cost on average.
>
> Using this patch, we compared the time of ksm_do_scan() function
> by adding kernel trace at the start-end position of operation.
> (ARM 32bit target android device,
> over 1000 sample time gap stamps average)
>
> On original KSM scan avg duration = 0.0166893 sec
> 24991.975619 : ksm_do_scan_start: START: ksm_do_scan
> 24991.990975 : ksm_do_scan_end: END: ksm_do_scan
> 24992.008989 : ksm_do_scan_start: START: ksm_do_scan
> 24992.016839 : ksm_do_scan_end: END: ksm_do_scan
> ...
>
> On patch KSM scan avg duration = 0.0041157 sec
> 41081.461312 : ksm_do_scan_start: START: ksm_do_scan
> 41081.466364 : ksm_do_scan_end: END: ksm_do_scan
> 41081.484767 : ksm_do_scan_start: START: ksm_do_scan
> 41081.487951 : ksm_do_scan_end: END: ksm_do_scan
> ...
>
> We have tested randomly so many times for the stability
> and couldn't see any abnormal issue until now.
> Also, we found out this patch can make some good advantage
> for the power consumption than KSM default enable.
>
> Signed-off-by: Kyeongdon Kim <kyeongdon.kim@xxxxxxx>
> ---
> mm/ksm.c | 49 +++++++++++++++++++++++++++++++++++++++++++++----
> 1 file changed, 45 insertions(+), 4 deletions(-)
>
> diff --git a/mm/ksm.c b/mm/ksm.c
> index be8f457..66ab4f4 100644
> --- a/mm/ksm.c
> +++ b/mm/ksm.c
> @@ -150,6 +150,7 @@ struct stable_node {
> struct hlist_head hlist;
> union {
> unsigned long kpfn;
> + u32 oldchecksum;
> unsigned long chain_prune_time;
> };
> /*

May be just checksum? i.e. that's can be "old", where checksum can change,
in stable tree, checksum also stable.

Also, as checksum are stable, may be that make a sense to move it out
of union? (I'm afraid of clashes)

Also, you miss update comment above struct stable_node, about checksum var.

Thanks for your comment, and we may change those lines like below :

+ * @oldchecksum: previous checksum of the page about a stable_node
 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
 */
Âstruct stable_node {
@@ -159,6 +160,7 @@ struct stable_node {
ÂÂÂÂÂÂÂÂ */
Â#define STABLE_NODE_CHAIN -1024
ÂÂÂÂÂÂÂ int rmap_hlist_len;
+ÂÂÂÂÂÂ u32 oldchecksum;
Â#ifdef CONFIG_NUMA

And I think if checksum are matched,
then we can use original memcmp logic in stable tree.
the worst case that I imagine is no page merging(just in that moment).
But, in my humble opinion, there will be no critical memory issue. but just return.
(as I said, we tested a lot to check some abnormal memory operation,
but so far, so good - only performance improvement)
> @@ -1522,7 +1523,7 @@ static __always_inline struct page *chain(struct stable_node **s_n_d,
> * This function returns the stable tree node of identical content if found,
> * NULL otherwise.
> */
> -static struct page *stable_tree_search(struct page *page)
> +static struct page *stable_tree_search(struct page *page, u32 checksum)
> {
> int nid;
> struct rb_root *root;
> @@ -1540,6 +1541,8 @@ static struct page *stable_tree_search(struct page *page)
>
> nid = get_kpfn_nid(page_to_pfn(page));
> root = root_stable_tree + nid;
> + if (!checksum)
> + return NULL;

That's not a pointer, and 0x0 - is a valid checksum.
Also, jhash2 not so collision free, i.e.:
jhash2((uint32_t *) &num, 2, 17);

Example of collisions, where hash = 0x0:
hash: 0x0 - num: 610041898
hash: 0x0 - num: 4893164379
hash: 0x0 - num: 16423540221
hash: 0x0 - num: 29036382188

You also compare values, so hash = 0, is a acceptable checksum.

well, if then, I can remove this check line.

Thanks,
anyway in general idea looks good.

Reviewed-by: Timofey Titovets <nefelim4ag@xxxxxxxxx>

--
Have a nice day,
Timofey.
Thanks a lot :)
Actually, our organization want to use this KSM feature in general,
but, current logic needs too high cost.
So I wish to change more light version.
Please kindly give your opinion on this idea.

Thanks,
Kyeongdon Kim