Bug #44916: client: syncfs flush is only fast with a single MDS - CephFS - Ceph

Actions

Copy link

Bug #44916

closed

client: syncfs flush is only fast with a single MDS

Added by Greg Farnum about 4 years ago. Updated 5 months ago.

Status:

Resolved

Priority:

Normal

Assignee:

Xiubo Li

Category:

Performance/Resource Usage

Target version:

Ceph - v19.0.0

% Done:

Source:

Tags:

backport_processed

Backport:

pacific

Regression:

Severity:

3 - minor

Reviewed:

Affected Versions:

ceph-qa-suite:

Component(FS):

Client

Labels (FS):

task(easy)

Pull request ID:

44255

Crash signature (v1):

Crash signature (v2):

Description

When we invoke Client::syncfs, we call into flush_caps_sync() and that invokes check_caps() for everything dirty, adding the internal flag CHECK_CAPS_SYNCHRONOUS to the last dirty cap in the list. check_caps() will set FLAG_SYNC on the MClientCaps message in that case.

But that method only sends the FLAG_SYNC on a single MDS session, and if we have dirty data for more than one MDS we will have to wait for all of them to decide to commit to disk! This may syncfs() quite slow on scale-out clusters.

Related issues 2 (0 open — 2 closed)