lustre-devel-lustre.org archive mirror
 help / color / mirror / Atom feed
From: James Simmons <jsimmons@infradead.org>
To: Andreas Dilger <adilger@whamcloud.com>,
	Oleg Drokin <green@whamcloud.com>, NeilBrown <neilb@suse.de>
Cc: Chris Horn <chris.horn@hpe.com>,
	Serguei Smirnov <ssmirnov@whamcloud.com>,
	Lustre Development List <lustre-devel@lists.lustre.org>
Subject: [lustre-devel] [PATCH 16/33] lnet: improve numeric NID to CPT hashing
Date: Sun,  2 Feb 2025 15:46:16 -0500	[thread overview]
Message-ID: <20250202204633.1148872-17-jsimmons@infradead.org> (raw)
In-Reply-To: <20250202204633.1148872-1-jsimmons@infradead.org>

From: Chris Horn <chris.horn@hpe.com>

Testing shows that the sum-by-multiplication of nid bytes method for
hashing a NID to a CPT does not have good distribution for gnilnd
and kfilnd NIDs. For these LNDs, use the old hash_long() method.

Also modify lustre_lnet_calc_cpt_of_nid() to ensure the specified
number of CPTs is greater than 0. This avoids a divide by zero.

HPE-bug-id: LUS-11576
WC-bug-id: https://jira.whamcloud.com/browse/LU-16797
Lustre-commit: bd12731f2f15b3903 ("LU-16797 lnet: improve numeric NID to CPT hashing")
Signed-off-by: Chris Horn <chris.horn@hpe.com>
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/50862
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Serguei Smirnov <ssmirnov@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
Signed-off-by: James Simmons <jsimmons@infradead.org>
---
 net/lnet/lnet/api-ni.c | 31 +++++++++++++++++++++----------
 1 file changed, 21 insertions(+), 10 deletions(-)

diff --git a/net/lnet/lnet/api-ni.c b/net/lnet/lnet/api-ni.c
index 585148b4b6aa..fa42cfd99aa7 100644
--- a/net/lnet/lnet/api-ni.c
+++ b/net/lnet/lnet/api-ni.c
@@ -1532,20 +1532,31 @@ static unsigned int
 lnet_nid4_cpt_hash(lnet_nid_t nid, unsigned int number)
 {
 	u64 key = nid;
-	u64 pair_bits = 0x0001000100010001LLU;
-	u64 mask = pair_bits * 0xFF;
-	u64 pair_sum;
+	u16 lnd = LNET_NETTYP(LNET_NIDNET(nid));
+	unsigned int cpt;
 
-	/* Use (sum-by-multiplication of nid bytes) mod (number of CPTs)
-	 * to match nid to a CPT.
-	 */
-	pair_sum = (key & mask) + ((key >> 8) & mask);
-	pair_sum = (pair_sum * pair_bits) >> 48;
+	if (lnd == KFILND || lnd == GNILND) {
+		cpt = hash_long(key, LNET_CPT_BITS);
+
+		/* NB: The number of CPTs needn't be a power of 2 */
+		if (cpt >= number)
+			cpt = (key + cpt + (cpt >> 1)) % number;
+	} else {
+		u64 pair_bits = 0x0001000100010001LLU;
+		u64 mask = pair_bits * 0xFF;
+		u64 pair_sum;
+		/* For ipv4 NIDs, use (sum-by-multiplication of nid bytes) mod
+		 * (number of CPTs) to match nid to a CPT.
+		 */
+		pair_sum = (key & mask) + ((key >> 8) & mask);
+		pair_sum = (pair_sum * pair_bits) >> 48;
+		cpt = (unsigned int)(pair_sum) % number;
+	}
 
 	CDEBUG(D_NET, "Match nid %s to cpt %u\n",
-	       libcfs_nid2str(nid), (unsigned int)(pair_sum) % number);
+	       libcfs_nid2str(nid), cpt);
 
-	return (unsigned int)(pair_sum) % number;
+	return cpt;
 }
 
 unsigned int
-- 
2.39.3

_______________________________________________
lustre-devel mailing list
lustre-devel@lists.lustre.org
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org

  parent reply	other threads:[~2025-02-02 20:54 UTC|newest]

Thread overview: 34+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-02-02 20:46 [lustre-devel] [PATCH 00/33] lustre: sync to OpenSFS branch May 31, 2023 James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 01/33] lnet: set msg field for lnet message header James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 02/33] Revert "lustre: llite: Check vmpage in releasepage" James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 03/33] lustre: llite: EIO is possible on a race with page reclaim James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 04/33] lustre: llite: add __GFP_NORETRY for read-ahead page James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 05/33] lustre: obd: change lmd flags to bitmap James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 06/33] lustre: uapi: cleanup FSFILT defines James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 07/33] lustre: obd: Reserve metadata overstriping flags James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 08/33] lnet: selftest: manage the workqueue state properly James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 09/33] lustre: remove cl_{offset, index, page_size} helpers James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 10/33] lustre: csdc: reserve layout bits for compress component James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 11/33] lustre: obd: replace simple_strtoul() James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 12/33] lnet: Use dynamic allocation for LND tunables James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 13/33] lustre: cksum: fix generating T10PI guard tags for partial brw page James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 14/33] lustre: llite: remove OBD_ -> CFS_ macros James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 15/33] lustre: obd: " James Simmons
2025-02-02 20:46 ` James Simmons [this message]
2025-02-02 20:46 ` [lustre-devel] [PATCH 17/33] lnet: libcfs: Remove unsed LASSERT_ATOMIC_* macros James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 18/33] lustre: misc: replace obsolete ioctl numbers James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 19/33] lustre: lmv: treat unknown hash type as sane type James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 20/33] lustre: llite: Fix return for non-queued aio James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 21/33] lnet: collect data about routes by using Netlink James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 22/33] lustre: ptlrpc: switch sptlrpc_rule_set_choose to large nid James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 23/33] lnet: use list_first_entry() where appropriate James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 24/33] lustre: statahead: using try lock for batched RPCs James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 25/33] lnet: libcfs: use round_up directly James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 26/33] lustre: mdc: md_open_data should keep ref on close_req James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 27/33] lustre: llite: update comment of ll_swap_layouts_close James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 28/33] lustre: ldlm: replace OBD_ -> CFS_ macros James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 29/33] lustre: mdc: remove " James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 30/33] lnet: libcfs: move cfs_expr_list_print to nidstrings.c James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 31/33] lnet: libcfs: Remove reference to LASSERT_ATOMIC_POS James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 32/33] lnet: ksocklnd: ksocklnd_ni_get_eth_intf_speed() must use only rtnl lock James Simmons
2025-02-02 20:46 ` [lustre-devel] [PATCH 33/33] lustre: ldlm: convert ldlm extent locks to linux extent-tree James Simmons

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20250202204633.1148872-17-jsimmons@infradead.org \
    --to=jsimmons@infradead.org \
    --cc=adilger@whamcloud.com \
    --cc=chris.horn@hpe.com \
    --cc=green@whamcloud.com \
    --cc=lustre-devel@lists.lustre.org \
    --cc=neilb@suse.de \
    --cc=ssmirnov@whamcloud.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).