curvefs: add a new distributed transaction model to improve rename performance

Signed-off-by: wanghai01 <seanhaizi@163.com>

Author: SeanHai

curvefs/conf/client.conf

@@ -85,6 +85,12 @@ fuseClient.getThreadPool=4
 # it gurantee the consistent of file after rename, otherwise you should
 # disable it for performance.
 fuseClient.enableMultiMountPointRename=true
+
+# the rename transaction models are different between version 1 and version 2
+# the v2 version greatly improves the performance of rename, especially in concurrent scenarios.
+# Node: v1 and v2 are incompatible and cannot be directly upgraded from a v1 cluster to v2.
+fuseClient.txVersion=1
+
 # splice will bring higher performance in some cases
 # but there might be a kernel issue that will cause kernel panic when enabling it
 # see https://lore.kernel.org/all/CAAmZXrsGg2xsP1CK+cbuEMumtrqdvD-NKnWzhNcvn71RV3c1yw@mail.gmail.com/

curvefs/conf/metaserver.conf

@@ -258,11 +258,17 @@ storage.rocksdb.unordered_write_buffer_size=67108864
 # for store inode which exclude its s3chunkinfo list (default: 3)
 storage.rocksdb.unordered_max_write_buffer_number=3
 # rocksdb column family's write_buffer_size
-# for store dentry and inode's s3chunkinfo list (unit: bytes, default: 128MB)
+# for store dentry and inode's s3chunkinfo list (unit: bytes, default: 64MB)
 storage.rocksdb.ordered_write_buffer_size=67108864
 # rocksdb column family's max_write_buffer_number
 # for store dentry and inode's s3chunkinfo list (default: 3)
 storage.rocksdb.ordered_max_write_buffer_number=3
+# rocksdb column family's write_buffer_size
+# for store tx lock and write (unit: bytes, default: 64MB)
+storage.rocksdb.tx_cf_write_buffer_size=67108864
+# rocksdb column family's max_write_buffer_number
+# for store tx lock and write (default: 3)
+storage.rocksdb.tx_cf_max_write_buffer_number=3
 # The target number of write history bytes to hold in memory (default: 20MB)
 storage.rocksdb.max_write_buffer_size_to_maintain=20971520
 # rocksdb memtable prefix bloom size ratio (size=write_buffer_size*memtable_prefix_bloom_size_ratio)
@@ -286,6 +292,8 @@ storage.rocksdb.perf_sampling_ratio=0
 # we will sending its with rpc streaming instead of
 # padding its into inode (default: 25000, about 25000 * 41 (byte) = 1MB)
 storage.s3_meta_inside_inode.limit_size=25000
+# TTL(millisecond) for tx lock
+storage.rocksdb.tx_lock_ttl_ms=5000
 
 # recycle options
 # metaserver scan recycle period, default 1h
@@ -317,7 +325,6 @@ excutorOpt.minRetryTimesForceTimeoutBackoff=5
 excutorOpt.maxRetryTimesBeforeConsiderSuspend=20
 # batch limit of get inode attr and xattr
 excutorOpt.batchInodeAttrLimit=10000
-excutorOpt.enableMultiMountPointRename=true
 
 #### metaCacheOpt
 # Gets the number of retries for the leader