Data-on-MDT phase II
(LU-10176)
|
|
| Status: | Resolved |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | Lustre 2.11.0 |
| Fix Version/s: | None |
| Type: | Technical task | Priority: | Minor |
| Reporter: | Andreas Dilger | Assignee: | Mikhail Pershin |
| Resolution: | Fixed | Votes: | 0 |
| Labels: | DoM | ||
| Issue Links: |
|
||||||||
| Rank (Obsolete): | 9223372036854775807 | ||||||||
| Description |
|
The SHORTIO functionality was landed in Lustre 2.11, at the same time as DoM, so these features do not currently work together. It may just be a matter of enabling the OBD_CONNECT_SHORTIO feature flag on the MDS connection, but it may be that more work is needed. One related possibility is to include DoM file’s dirty data in the close RPC (or conversely a close in the write RPC, whichever makes more sense). That would further reduce the RPC count for small files. Also, since the DoM close will also be updating attributes like atime or mtime/ctime, we may as well write the data at the same time to avoid more disk IO if they are done in separate RPCs. |
| Comments |
| Comment by Patrick Farrell (Inactive) [ 28/Nov/17 ] |
|
Hm. I... I think we didn't add SHORT_IO to the OST connection flags (server, not client side)? I'm confused. I clearly remember manually verifying it was working, by noting the lack of an RDMA. Am I wrong? Did we not really enable it? I'll try to test a bit later. |
| Comment by Patrick Farrell (Inactive) [ 28/Nov/17 ] |
|
Uck. Short i/o doesn't seem to be working at all, even with the flags added. Not sure what changed from when I tested it carefully a while back, but something did on the way to landing, it seems. I'll try to find time later to investigate. |
| Comment by Patrick Farrell (Inactive) [ 28/Nov/17 ] |
|
Mike, does this mean you're going to figure out if short i/o is working on the OST side too? I haven't dug in yet. |
| Comment by Andreas Dilger [ 29/Nov/17 ] |
|
I see OBD_CONNECT_SHORTIO is included in ocd_connect_flags for the OST, but not for the MDT ocd_connect_flags. However, I do NOT see OBD_CONNECT_SHORTIO in OST_CONNECT_SUPPORTED or MDT_CONNECT_SUPPORTED. Maybe you missed to add it to the patch, but it was on the server that you tested against? |
| Comment by Patrick Farrell (Inactive) [ 29/Nov/17 ] |
|
Must be, yeah. Uck! I also added it to OST_CONNECT_SUPPORTED real quick and still didn't get short i/o yesterday. Not sure why and not sure when I can dig in. I'll try to find time soon. |
| Comment by Patrick Farrell (Inactive) [ 07/Dec/17 ] |
|
I actually did this in: Which I marked against |
| Comment by Joseph Gmitter (Inactive) [ 07/Mar/18 ] |
|
https://review.whamcloud.com/#/c/30435/ ( |
| Comment by Patrick Farrell (Inactive) [ 07/Mar/18 ] |
|
Agreed, sorry I didn't get it earlier. |
| Comment by Patrick Farrell (Inactive) [ 07/Mar/18 ] |
|
Resolved in |
| Comment by Joseph Gmitter (Inactive) [ 07/Mar/18 ] |
|
Thanks Patrick. |