cifs: deal with the channel loading lag while picking channels
Subsystems (3)
PATCH 2/2 cifs: deal with the channel loading lag while picking channels
From: Shyam Prasad N <[email protected]>
Our current approach to select a channel for sending requests is this:
1. iterate all channels to find the min and max queue depth
2. if min and max are not the same, pick the channel with min depth
3. if min and max are same, round robin, as all channels are equally loaded
The problem with this approach is that there's a lag between selecting a channel and sending the request (that increases the queue depth on the channel). While these numbers will eventually catch up, there could be a skew in the channel usage, depending on the application's I/O parallelism and the server's speed of handling requests.
With sufficient parallelism, this lag can artificially increase the queue depth, thereby impacting the performance negatively.
This change will change the step 1 above to start the iteration from the last selected channel. This is to reduce the skew in channel usage even in the presence of this lag.
--- fs/smb/client/transport.c | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-)
fs/smb/client/transport.c +7 -7
@@ -1015,14 +1015,16 @@ struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)uint index =0 ;unsigned int min_in_flight =UINT_MAX , max_in_flight =0 ;struct TCP_Server_Info *server =NULL ; -int i; +int i, start, cur;if (!ses)return NULL ;spin_lock (&ses->chan_lock); + start =atomic_inc_return (&ses->chan_seq);for (i =0 ; i < ses->chan_count; i++) { - server = ses->chans[i].server; + cur = (start + i) % ses->chan_count; + server = ses->chans[cur].server;if (!server || server->terminate)continue ;
@@ -1039,17 +1041,15 @@ struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)*/ if (server->in_flight < min_in_flight) { min_in_flight = server->in_flight; - index = i; + index = cur; }if (server->in_flight > max_in_flight) max_in_flight = server->in_flight; }/* if all channels are equally loaded, fall back to round-robin */ -if (min_in_flight == max_in_flight) { - index = (uint )atomic_inc_return (&ses->chan_seq); - index %= ses->chan_count; - } +if (min_in_flight == max_in_flight) + index = (uint )start % ses->chan_count; server = ses->chans[index].server;spin_unlock (&ses->chan_lock);
Signature · 1 line
2.43.0
tentatively merged into cifs-2.6.git for-next pending more reviews and testing
Quoted 75 lines · On Wed, Feb 12, 2025 at 1:35 AM <[email protected]> wrote:
From: Shyam Prasad N <[email protected]>
Our current approach to select a channel for sending requests is this:
1. iterate all channels to find the min and max queue depth
2. if min and max are not the same, pick the channel with min depth
3. if min and max are same, round robin, as all channels are equally loadedThe problem with this approach is that there's a lag between selecting a channel and sending the request (that increases the queue depth on the channel). While these numbers will eventually catch up, there could be a skew in the channel usage, depending on the application's I/O parallelism and the server's speed of handling requests.
With sufficient parallelism, this lag can artificially increase the queue depth, thereby impacting the performance negatively.
This change will change the step 1 above to start the iteration from the last selected channel. This is to reduce the skew in channel usage even in the presence of this lag.
Fixes ea90708d3cf3 ("cifs: use the least loaded channel for sending requests")Cc <[email protected]>Signed-off-by Shyam Prasad N <[email protected]>--- fs/smb/client/transport.c | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-)fs/smb/client/transport.c +7
-8@@ -1015,14 +1015,16 @@ struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)uint index =0 ;unsigned int min_in_flight =UINT_MAX , max_in_flight =0 ;struct TCP_Server_Info *server =NULL ; -int i; +int i, start, cur;if (!ses)return NULL ;spin_lock (&ses->chan_lock); + start =atomic_inc_return (&ses->chan_seq);for (i =0 ; i < ses->chan_count; i++) { - server = ses->chans[i].server; + cur = (start + i) % ses->chan_count; + server = ses->chans[cur].server;if (!server || server->terminate)continue ;@@ -1039,17 +1041,15 @@ struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)*/ if (server->in_flight < min_in_flight) { min_in_flight = server->in_flight; - index = i; + index = cur; }if (server->in_flight > max_in_flight) max_in_flight = server->in_flight; }/* if all channels are equally loaded, fall back to round-robin */ -if (min_in_flight == max_in_flight) { - index = (uint )atomic_inc_return (&ses->chan_seq); - index %= ses->chan_count; - } +if (min_in_flight == max_in_flight) + index = (uint )start % ses->chan_count; server = ses->chans[index].server;spin_unlock (&ses->chan_lock); --2.43.0
Signature · 3 lines
Thanks, Steve
Shyam,
Could the two test failures in:
http://smb311-linux-testing.southcentralus.cloudapp.azure.com/#/builders/3/builds/386
be related to this patch ie tests generic/133 and generic/471 failing with hundreds of "CIFS: Status code returned 0xc000009a STATUS_INSUFFICIENT_RESOURCES" errors. The only failing tests were multichannel related.
This test run was with these six patches:600ed21fe802 (HEAD -> for-next, origin/for-next, origin/HEAD) netfs:
Fix a number of read-retry hangs9f75ff5536b1 smb: client, common: Avoid multiple -Wflex-array-member-not-at-end warningsfab0eddb9fe7 cifs: Treat unhandled directory name surrogate reparse points as mount directory nodes69476da76b9c cifs: Throw -EOPNOTSUPP error on unsupported reparse point type from parse_reparse_point()ef590eae88cf cifs: deal with the channel loading lag while picking channels f1bf10d7e909 cifs: pick channels for individual subrequests
Anyone else seeing the same errors with multichannel on these tests?
be related
Quoted 85 lines · Steve French
tentatively merged into cifs-2.6.git for-next pending more reviews and testing
Quoted 73 lines · On Wed, Feb 12, 2025 at 1:35 AM <[email protected]> wrote:
From: Shyam Prasad N <[email protected]>
Our current approach to select a channel for sending requests is this:
1. iterate all channels to find the min and max queue depth
2. if min and max are not the same, pick the channel with min depth
3. if min and max are same, round robin, as all channels are equally loadedThe problem with this approach is that there's a lag between selecting a channel and sending the request (that increases the queue depth on the channel). While these numbers will eventually catch up, there could be a skew in the channel usage, depending on the application's I/O parallelism and the server's speed of handling requests.
With sufficient parallelism, this lag can artificially increase the queue depth, thereby impacting the performance negatively.
This change will change the step 1 above to start the iteration from the last selected channel. This is to reduce the skew in channel usage even in the presence of this lag.
Fixes ea90708d3cf3 ("cifs: use the least loaded channel for sending requests")Cc <[email protected]>Signed-off-by Shyam Prasad N <[email protected]>--- fs/smb/client/transport.c | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-)fs/smb/client/transport.c +7
-8@@ -1015,14 +1015,16 @@ struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)uint index =0 ;unsigned int min_in_flight =UINT_MAX , max_in_flight =0 ;struct TCP_Server_Info *server =NULL ; -int i; +int i, start, cur;if (!ses)return NULL ;spin_lock (&ses->chan_lock); + start =atomic_inc_return (&ses->chan_seq);for (i =0 ; i < ses->chan_count; i++) { - server = ses->chans[i].server; + cur = (start + i) % ses->chan_count; + server = ses->chans[cur].server;if (!server || server->terminate)continue ;@@ -1039,17 +1041,15 @@ struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)*/ if (server->in_flight < min_in_flight) { min_in_flight = server->in_flight; - index = i; + index = cur; }if (server->in_flight > max_in_flight) max_in_flight = server->in_flight; }/* if all channels are equally loaded, fall back to round-robin */ -if (min_in_flight == max_in_flight) { - index = (uint )atomic_inc_return (&ses->chan_seq); - index %= ses->chan_count; - } +if (min_in_flight == max_in_flight) + index = (uint )start % ses->chan_count; server = ses->chans[index].server;spin_unlock (&ses->chan_lock); --Quoted 2 lines
2.43.0
--
Thanks,Steve
Signature · 3 lines
Thanks, Steve