defer gc during block processing (#3384)

arnetheduck · web-flow · commit 3a009158d3e3 · 2025-06-20T09:59:51.000+02:00
During block processing, we allocate lots and lots of small objects
(VertexRef and friends) which causes overhead as the GC is run without
much benefit - this can be up to 15% of CPU time during block import
when running with a large vertex cache.
diff --git a/execution_chain/core/executor/process_block.nim b/execution_chain/core/executor/process_block.nim
@@ -280,15 +280,21 @@ proc processBlock*(
     taskpool: Taskpool = nil,
 ): Result[void, string] =
   ## Generalised function to processes `blk` for any network.
-  ?vmState.procBlkPreamble(blk, skipValidation, skipReceipts, skipUncles, taskpool)
 
-  # EIP-3675: no reward for miner in POA/POS
-  if not vmState.com.proofOfStake(blk.header, vmState.ledger.txFrame):
-    vmState.calculateReward(blk.header, blk.uncles)
+  # Processing a block involves making lots and lots of small memory allocations
+  # meaning that GC overhead can make up for 15% of processing time in extreme
+  # cases - since each block is bounded in the amount of memory needed, we can
+  # run collection once per block instead.
+  deferGc:
+    ?vmState.procBlkPreamble(blk, skipValidation, skipReceipts, skipUncles, taskpool)
 
-  ?vmState.procBlkEpilogue(blk, skipValidation, skipReceipts, skipStateRootCheck)
+    # EIP-3675: no reward for miner in POA/POS
+    if not vmState.com.proofOfStake(blk.header, vmState.ledger.txFrame):
+      vmState.calculateReward(blk.header, blk.uncles)
 
-  ok()
+    ?vmState.procBlkEpilogue(blk, skipValidation, skipReceipts, skipStateRootCheck)
+
+    ok()
 
 # ------------------------------------------------------------------------------
 # End
diff --git a/execution_chain/utils/utils.nim b/execution_chain/utils/utils.nim
@@ -173,3 +173,18 @@ func weiAmount*(w: Withdrawal): UInt256 =
 func isGenesis*(header: Header): bool =
   header.number == 0'u64 and
     header.parentHash == GENESIS_PARENT_HASH
+
+template deferGc*(body: untyped): untyped =
+  when declared(GC_disable):
+    GC_disable()
+
+  when declared(GC_enable):
+    defer:
+      GC_enable()
+      # Perform a small allocation which indirectly runs the garbage collector -
+      # unlike GC_fullCollect, this will use the usual nim heuristic for running
+      # the cycle colllector (which would be extremely expensive to run on each
+      # collection)
+      discard newSeq[int](1)
+
+  body
