private void TrimCacheIfOverLimit()
{
if (Interlocked.Read(ref _cacheSize) < _maxSize)
{
return;
}
IList <(KeyValuePair <ulong, SegmentCacheEntry> CacheEntry, int LastAccessTimestamp)> entries = SnapshotNonMinSizeCacheItems();;
if (entries.Count == 0)
{
return;
}
// Try to cut ourselves down to about 85% of our max capaity, otherwise just hang out right at that boundary and the next entry we add we end up having to
// scavenge again, and again, and again...
uint requiredCutAmount = (uint)(_maxSize * 0.15);
long desiredSize = (long)(_maxSize * 0.85);
bool haveUpdatedSnapshot = false;
uint cutAmount = 0;
while (cutAmount < requiredCutAmount)
{
// We could also be trimming on other threads, so if collectively we have brought ourselves below 85% of our max capacity then we are done
if (Interlocked.Read(ref _cacheSize) <= desiredSize)
{
break;
}
// find the largest item of the 10% of least recently accessed (remaining) items
uint largestSizeSeen = 0;
int curItemIndex = (entries.Count - 1) - (int)(entries.Count * 0.10);
if (curItemIndex < 0)
{
return;
}
int removalTargetIndex = -1;
while (curItemIndex < entries.Count)
{
KeyValuePair <ulong, SegmentCacheEntry> curItem = entries[curItemIndex].CacheEntry;
// TODO: CurrentSize is constantly in flux, other threads are paging data in and out, so while I can read it in the below check there is no guarantee that by the time I get
// to the code that tries to REMOVE this item that CurrentSize hasn't changed
//
// >= so we prefer the largest item that is least recently accessed, ensuring we don't remove a segment that is being actively modified now (should
// never happen since we also update that segments accessed timestamp, but, defense in depth).
if ((curItem.Value.CurrentSize - curItem.Value.MinSize) >= largestSizeSeen)
{
largestSizeSeen = (uint)(curItem.Value.CurrentSize - curItem.Value.MinSize);
removalTargetIndex = curItemIndex;
}
curItemIndex++;
}
if (removalTargetIndex < 0)
{
// If we are already below our desired size or we have already re-snapshotted one time, then just give up
if (haveUpdatedSnapshot || (Interlocked.Read(ref _cacheSize) <= desiredSize))
{
break;
}
// we failed to find any non MinSize entris in the last 10% of entries. Since our snapshot originally ONLY contained non-MinSize entries this likely
// means other threads are also trimming. Re-snapshot and try again.
entries = SnapshotNonMinSizeCacheItems();
haveUpdatedSnapshot = true;
if (entries.Count == 0)
{
break;
}
continue;
}
SegmentCacheEntry targetItem = entries[removalTargetIndex].CacheEntry.Value;
if (HeapSegmentCacheEventSource.Instance.IsEnabled())
{
HeapSegmentCacheEventSource.Instance.PageOutDataStart();
}
long sizeRemoved = targetItem.PageOutData();
if (HeapSegmentCacheEventSource.Instance.IsEnabled())
{
HeapSegmentCacheEventSource.Instance.PageOutDataEnd(sizeRemoved);
}
// Whether or not we managed to remove any memory for this item (another thread may have removed it all before we could), remove it from our list of
// entries to consider
entries.RemoveAt(removalTargetIndex);
if (sizeRemoved != 0)
{
Interlocked.Add(ref _cacheSize, -sizeRemoved);
cutAmount += (uint)sizeRemoved;
}
}
}
private void TrimCacheIfOverLimit(ulong modifiedSegmentAddress)
{
if (Interlocked.Read(ref _cacheSize) < _maxSize)
{
return;
}
// Select all cache entries which aren't at their min-size
//
// NOTE: We snapshot the LastAccessTickCount values here because there is the case where the Sort function will throw an exception if it tests two entries and the
// lhs rhs comparison is inconsistent when reveresed (i.e. something like lhs < rhs is true but then rhs < lhs is also true). This sound illogical BUT it can happen
// if between the two comparisons the LastAccessTickCount changes (because other threads are concurrently accessing these same entries), in that case we would trigger
// this exception, which is bad :)
IEnumerable <(KeyValuePair <ulong, SegmentCacheEntry> CacheEntry, long LastAccessTickCount)>?items = null;
List <(KeyValuePair <ulong, SegmentCacheEntry> CacheEntry, long LastAccessTickCount)>? entries = null;
_cacheLock.EnterReadLock();
try
{
items = _cache.Where((kvp) => kvp.Value.CurrentSize != kvp.Value.MinSize).Select((kvp) => (CacheEntry: kvp, kvp.Value.LastAccessTickCount));
entries = new List <(KeyValuePair <ulong, SegmentCacheEntry> CacheEntry, long LastAccessTickCount)>(items);
}
finally
{
_cacheLock.ExitReadLock();
}
// Flip the sort order to the LEAST recently accessed items (i.e. the ones whose LastAccessTickCount are furthest in history) end up at the END of the array,
//
// NOTE: Using tickcounts is succeptible to roll-over, but worst case scenario we remove a slightly more recently used one thinking it is older, not a huge deal
// and using DateTime.Now to get a non-roll-over succeptible timestamp showed up as 5% of scenario time in PerfView :(
entries.Sort((lhs, rhs) => rhs.LastAccessTickCount.CompareTo(lhs.LastAccessTickCount));
// Try to cut ourselves down to about 85% of our max capaity, otherwise just hang out right at that boundary and the next entry we add we end up having to
// scavenge again, and again, and again...
uint requiredCutAmount = (uint)(_maxSize * 0.15);
long desiredSize = (long)(_maxSize * 0.85);
uint cutAmount = 0;
while (cutAmount < requiredCutAmount)
{
// We could also be trimming on other threads, so if collectively we have brought ourselves below 85% of our max capacity then we are done
if (Interlocked.Read(ref _cacheSize) <= desiredSize)
{
break;
}
// find the largest item of the 10% of least recently accessed (remaining) items
uint largestSizeSeen = 0;
int curItemIndex = (entries.Count - 1) - (int)(entries.Count * 0.10);
if (curItemIndex < 0)
{
return;
}
int removalTargetIndex = -1;
while (curItemIndex < entries.Count)
{
KeyValuePair <ulong, SegmentCacheEntry> curItem = entries[curItemIndex].CacheEntry;
// >= so we prefer the largest item that is least recently accessed, ensuring we don't remove a segment that is being actively modified now (should
// never happen since we also update that segments accessed timestamp, but, defense in depth).
//
// NOTE: We subtract MinSize from CurrentSize assuming the cache will be able to page out its data and thus we won't
// actually remove its entry below. If not we will correct this value later in the 'we remove the entire cache entry'
// block (the !PageOutData block below).
if ((curItem.Value.CurrentSize - curItem.Value.MinSize) >= largestSizeSeen && (curItem.Key != modifiedSegmentAddress))
{
largestSizeSeen = (uint)(curItem.Value.CurrentSize - curItem.Value.MinSize);
removalTargetIndex = curItemIndex;
}
curItemIndex++;
}
if (removalTargetIndex == -1)
{
return;
}
SegmentCacheEntry targetItem = entries[removalTargetIndex].CacheEntry.Value;
if (HeapSegmentCacheEventSource.Instance.IsEnabled())
{
HeapSegmentCacheEventSource.Instance.PageOutDataStart();
}
long removedBytes = targetItem.PageOutData();
if (HeapSegmentCacheEventSource.Instance.IsEnabled())
{
HeapSegmentCacheEventSource.Instance.PageOutDataEnd(removedBytes);
}
// Whether or not we managed to remove any memory for this item (another thread may have removed it all before we could), remove it from our list of
// entries to consider
entries.RemoveAt(removalTargetIndex);
if (removedBytes != 0)
{
Interlocked.Add(ref _cacheSize, -removedBytes);
cutAmount += (uint)removedBytes;
}
}
}