[arvados] updated: 2.7.0-6012-g222b45df0a

[git repository hosting]

Summary of changes:
 sdk/go/arvados/fs_collection.go | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)

       via  222b45df0a30dad5189599d7942271c63f46f72d (commit)
      from  1dcde0921ddb62ae1a4ea01bc9d5179c6a994882 (commit)

Those revisions listed above that are new to this repository have
not appeared on any other notification email; so we list those
revisions in full, below.


commit 222b45df0a30dad5189599d7942271c63f46f72d
Author: Tom Clegg <tom at curii.com>
Date:   Mon Feb 26 16:28:07 2024 -0500

    18961: Hold off multiple pre-fetch for 10 seconds, not forever.
    
    In the keep-web usage pattern, the "don't pre-fetch this block again"
    flag prevents the prefetch mechanism from working more than once on a
    collectionfs that is kept in memory for a long time and used
    repeatedly.
    
    Arvados-DCO-1.1-Signed-off-by: Tom Clegg <tom at curii.com>

diff --git a/sdk/go/arvados/fs_collection.go b/sdk/go/arvados/fs_collection.go
index c3a210faa3..875a942210 100644
--- a/sdk/go/arvados/fs_collection.go
+++ b/sdk/go/arvados/fs_collection.go
@@ -645,7 +645,7 @@ func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr
 	}
 	if ss, ok := fn.segments[ptr.segmentIdx].(storedSegment); ok {
 		ss.locator = fn.fs.refreshSignature(ss.locator)
-		ss.didRead = true
+		ss.didRead = time.Now()
 		fn.segments[ptr.segmentIdx] = ss
 	}
 
@@ -1227,6 +1227,10 @@ var (
 	// At runtime this is a no-op. Test cases replace this func
 	// with a call to Add(1).
 	profAdd1 = func(*atomic.Int64) {}
+
+	// Don't pre-fetch a block if we already tried to read it this
+	// recently.
+	prefetchHoldDuration = 10 * time.Second
 )
 
 // Prefetch file data based on expected future usage.
@@ -1284,12 +1288,12 @@ func (dn *dirnode) prefetch(fn *filenode, name string, ptr filenodePtr) {
 		}
 		if next, ok := fn.segments[inext].(storedSegment); !ok {
 			todo -= fn.segments[inext].Len()
-		} else if next.didRead {
+		} else if !next.didRead.IsZero() && time.Since(next.didRead) < prefetchHoldDuration {
 			todo -= next.size
 			lastlocator = next.locator
 		} else {
 			profAdd1(&prefetchReadCurrent)
-			next.didRead = true
+			next.didRead = time.Now()
 			fn.segments[inext] = next
 			go next.ReadAt([]byte{}, 0)
 			todo -= next.size
@@ -1340,13 +1344,13 @@ func (dn *dirnode) prefetch(fn *filenode, name string, ptr filenodePtr) {
 			if next.locator == lastlocator {
 				// we already subtracted this block's
 				// size from todo
-			} else if next.didRead {
+			} else if !next.didRead.IsZero() && time.Since(next.didRead) < prefetchHoldDuration {
 				// someone already fetched this
 				// segment
 				todo -= next.size
 			} else {
 				profAdd1(&prefetchReadNext)
-				next.didRead = true
+				next.didRead = time.Now()
 				fn.segments[inext] = next
 				go next.ReadAt([]byte{}, 0)
 				todo -= next.size
@@ -1937,10 +1941,10 @@ type storedSegment struct {
 	offset  int // position of segment within the stored block
 	length  int // bytes in this segment (offset + length <= size)
 
-	// set when we first try to read from this segment, and
+	// Set when we first try to read from this segment, and
 	// checked before pre-fetch, to avoid unnecessary cache
-	// thrashing
-	didRead bool
+	// thrashing. See prefetchHoldDuration.
+	didRead time.Time
 }
 
 func (se storedSegment) Len() int {

-----------------------------------------------------------------------


hooks/post-receive
--