diff options
Diffstat (limited to 'drivers/staging/lustre/sysfs-fs-lustre')
-rw-r--r-- | drivers/staging/lustre/sysfs-fs-lustre | 103 |
1 files changed, 103 insertions, 0 deletions
diff --git a/drivers/staging/lustre/sysfs-fs-lustre b/drivers/staging/lustre/sysfs-fs-lustre index 1e302e8..873e2cf 100644 --- a/drivers/staging/lustre/sysfs-fs-lustre +++ b/drivers/staging/lustre/sysfs-fs-lustre @@ -40,6 +40,109 @@ Description: e.g. dd.1253 nodelocal - use jobid_name value from above. +What: /sys/fs/lustre/timeout +Date: June 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls "lustre timeout" variable, also known as obd_timeout + in some old manual. In the past obd_timeout was of paramount + importance as the timeout value used everywhere and where + other timeouts were derived from. These days it's much less + important as network timeouts are mostly determined by + AT (adaptive timeouts). + Unit: seconds, default: 100 + +What: /sys/fs/lustre/max_dirty_mb +Date: June 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls total number of dirty cache (in megabytes) allowed + across all mounted lustre filesystems. + Since writeout of dirty pages in Lustre is somewhat expensive, + when you allow to many dirty pages, this might lead to + performance degradations as kernel tries to desperately + find some pages to free/writeout. + Default 1/2 RAM. Min value 4, max value 9/10 of RAM. + +What: /sys/fs/lustre/debug_peer_on_timeout +Date: June 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Control if lnet debug information should be printed when + an RPC timeout occurs. + 0 disabled (default) + 1 enabled + +What: /sys/fs/lustre/dump_on_timeout +Date: June 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls if Lustre debug log should be dumped when an RPC + timeout occurs. This is useful if yout debug buffer typically + rolls over by the time you notice RPC timeouts. + +What: /sys/fs/lustre/dump_on_eviction +Date: June 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls if Lustre debug log should be dumped when an this + client is evicted from one of the servers. + This is useful if yout debug buffer typically rolls over + by the time you notice the eviction event. + +What: /sys/fs/lustre/at_min +Date: July 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls minimum adaptive timeout in seconds. If you encounter + a case where clients timeout due to server-reported processing + time being too short, you might consider increasing this value. + One common case of this if the underlying network has + unpredictable long delays. + Default: 0 + +What: /sys/fs/lustre/at_max +Date: July 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls maximum adaptive timeout in seconds. If at_max timeout + is reached for an RPC, the RPC will time out. + Some genuinuely slow network hardware might warrant increasing + this value. + Setting this value to 0 disables Adaptive Timeouts + functionality and old-style obd_timeout value is then used. + Default: 600 + +What: /sys/fs/lustre/at_extra +Date: July 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls how much extra time to request for unfinished requests + in processing in seconds. Normally a server-side parameter, it + is also used on the client for responses to various LDLM ASTs + that are handled with a special server thread on the client. + This is a way for the servers to ask the clients not to time + out the request that reached current servicing time estimate + yet and give it some more time. + Default: 30 + +What: /sys/fs/lustre/at_early_margin +Date: July 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls when to send the early reply for requests that are + about to timeout as an offset to the estimated service time in + seconds.. + Default: 5 + +What: /sys/fs/lustre/at_history +Date: July 2015 +Contact: "Oleg Drokin" <oleg.drokin@intel.com> +Description: + Controls for how many seconds to remember slowest events + encountered by adaptive timeouts code. + Default: 600 + What: /sys/fs/lustre/llite/<fsname>-<uuid>/blocksize Date: May 2015 Contact: "Oleg Drokin" <oleg.drokin@intel.com> |