[RESP] Optimize command parsing with SIMD fast path and O(1) hash table lookup

.github/copilot-instructions.md

@@ -92,7 +92,7 @@ Full guide: https://microsoft.github.io/garnet/docs/dev/garnet-api
 ### Steps for a new built-in command:
 
 1. **Define the command**: Add enum value to `RespCommand` in `libs/server/Resp/Parser/RespCommand.cs`. For object commands (List, SortedSet, Hash, Set), also add a value to the `[ObjectName]Operation` enum in `libs/server/Objects/[ObjectName]/[ObjectName]Object.cs`.
-2. **Add parsing logic**: In `libs/server/Resp/Parser/RespCommand.cs`, add to `FastParseCommand` (fixed arg count) or `FastParseArrayCommand` (variable args).
+2. **Add parsing logic**: In `libs/server/Resp/Parser/RespCommandHashLookup.cs`, add an entry to `PopulatePrimaryTable()` (e.g., `Add("MYNEWCMD", RespCommand.MYNEWCMD)`). For commands with subcommands, set `hasSub: true` and add a subcommand table. The hash table provides O(1) lookup for all command name lengths.
 3. **Declare the API method**: Add method signature to `IGarnetReadApi` (read-only) or `IGarnetApi` (read-write) in `libs/server/API/IGarnetApi.cs`.
 4. **Implement the network handler**: Add a method to `RespServerSession` (the class is split across ~22 partial `.cs` files — object commands go in `libs/server/Resp/Objects/[ObjectName]Commands.cs`, others in `libs/server/Resp/BasicCommands.cs`, `ArrayCommands.cs`, `AdminCommands.cs`, `KeyAdminCommands.cs`, etc.). The handler parses arguments from the network buffer via `parseState.GetArgSliceByRef(i)` (returns `ref PinnedSpanByte`), calls the storage API, and writes the RESP response using `RespWriteUtils` helper methods, then calls `SendAndReset()` to flush the response buffer.
 5. **Add dispatch route**: In `libs/server/Resp/RespServerSession.cs`, add a case to `ProcessBasicCommands` or `ProcessArrayCommands` calling the handler from step 4.

.github/skills/add-garnet-command/SKILL.md

@@ -82,17 +82,39 @@ EVALSHA, // Note: Update LastDataCommand if adding new data commands after this
 
 **File:** `libs/server/Resp/Parser/RespCommand.cs`
 
-Two parsing paths exist:
+Two parsing tiers exist:
 
-### Fast path: `FastParseCommand()` / `FastParseArrayCommand()`
-Two fast-path methods exist with different constraints:
-- **`FastParseCommand()`**: For commands with a fixed number of arguments and command names up to **9 characters**. Uses `ulong` pointer comparisons on `(count << 4) | length` patterns.
-- **`FastParseArrayCommand()`**: For commands with a variable number of arguments and command names up to **16 characters**. Uses similar `ulong` comparison patterns but accommodates longer names.
+### Hash table path: `RespCommandHashLookup` (primary path for most commands)
+The hash table in `libs/server/Resp/Parser/RespCommandHashLookup.cs` provides O(1) lookup for all built-in commands. This is the **recommended path for all new commands**.
 
-Only add here if the command name is a simple word (no dots or special characters).
+**To add a new primary command**, add one line to `PopulatePrimaryTable()`:
+```csharp
+Add("DELIFGREATER", RespCommand.DELIFGREATER);
+```
 
-### Slow path: `SlowParseCommand()`
-For longer names, dot-prefixed names (like `RI.CREATE`), or names that don't fit the fast-path pattern.
+**To add a command with subcommands** (e.g., `MYPARENT SUBCMD`):
+1. Add the parent with `hasSub: true`:
+   ```csharp
+   Add("MYPARENT", RespCommand.MYPARENT, hasSub: true);
+   ```
+2. Define the subcommand array:
+   ```csharp
+   private static readonly (string Name, RespCommand Command)[] MyparentSubcommands =
+   [
+       ("SUBCMD1", RespCommand.MYPARENT_SUBCMD1),
+       ("SUBCMD2", RespCommand.MYPARENT_SUBCMD2),
+   ];
+   ```
+3. Build the table in the static constructor:
+   ```csharp
+   myparentSubTable = BuildSubTable(MyparentSubcommands, out myparentSubTableMask);
+   ```
+4. Wire it into `LookupSubcommand()`:
+   ```csharp
+   RespCommand.MYPARENT => (myparentSubTable, myparentSubTableMask),
+   ```
+
+**⚠️ Important:** Use the exact wire-protocol spelling for the hash table name string. Some commands use hyphens (e.g., `"SET-CONFIG-EPOCH"` not `"SETCONFIGEPOCH"`). Check `CmdStrings.cs` for the canonical spelling.
 
 **⚠️ Convention:** Define the command name string in **`libs/server/Resp/CmdStrings.cs`** and reference it from the parser, rather than using inline `"..."u8` literals. This keeps command name strings centralized and reusable (e.g., for error messages).
 
@@ -101,26 +123,13 @@ For longer names, dot-prefixed names (like `RI.CREATE`), or names that don't fit
 public static ReadOnlySpan<byte> DELIFGREATER => "DELIFGREATER"u8;
 ```
 
-**Pattern for slow-path commands:**
-```csharp
-else if (command.SequenceEqual(CmdStrings.DELIFGREATER))
-{
-    return RespCommand.DELIFGREATER;
-}
-```
-
-**Pattern for dot-prefixed commands (e.g., `RI.CREATE`):**
-```csharp
-else if (command.SequenceEqual(CmdStrings.RICREATE))
-{
-    return RespCommand.RICREATE;
-}
-```
-
-Add this before the final `return RespCommand.INVALID;` at the end of `SlowParseCommand`.
+### SIMD fast path: `FastParseCommand()` (optional, for hottest commands only)
+Static `Vector128<byte>` patterns that match the full RESP encoding (`*N\r\n$L\r\nCMD\r\n`) in a single 16-byte comparison. Only needed for the most performance-critical commands with:
+- Fixed argument count
+- Command names of 3-6 characters
+- No dots or special characters
 
-**⚠️ Caveat: Dot-prefixed commands and ACL**
-If your command uses a dot (e.g., `RI.CREATE`), you must also update **`libs/server/ACL/ACLParser.cs`** so that the ACL system can map the dotted wire name to the enum name. Search for how existing dot-handling works (look for `Replace(".", "")` or similar normalization).
+Most new commands should **NOT** be added here — the hash table + MRU cache provide excellent performance for all commands. Only add SIMD patterns if benchmarking shows the command is a bottleneck.
 
 ---
 

benchmark/BDN.benchmark/Operations/CommandParsingBenchmark.cs

@@ -0,0 +1,242 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT license.
+
+using BenchmarkDotNet.Attributes;
+using Garnet.server;
+
+namespace BDN.benchmark.Operations
+{
+    /// <summary>
+    /// Benchmark for RESP command parsing only (no storage operations).
+    /// Calls ParseRespCommandBuffer directly to measure pure parsing throughput
+    /// across all optimization tiers.
+    /// </summary>
+    [MemoryDiagnoser]
+    public unsafe class CommandParsingBenchmark : OperationsBase
+    {
+        // Tier 1a: SIMD Vector128 fast path (3-6 char commands with fixed arg counts)
+        static ReadOnlySpan<byte> CMD_PING => "*1\r\n$4\r\nPING\r\n"u8;
+        static ReadOnlySpan<byte> CMD_GET => "*2\r\n$3\r\nGET\r\n$1\r\na\r\n"u8;
+        static ReadOnlySpan<byte> CMD_SET => "*3\r\n$3\r\nSET\r\n$1\r\na\r\n$1\r\nb\r\n"u8;
+        static ReadOnlySpan<byte> CMD_INCR => "*2\r\n$4\r\nINCR\r\n$1\r\ni\r\n"u8;
+        static ReadOnlySpan<byte> CMD_EXISTS => "*2\r\n$6\r\nEXISTS\r\n$1\r\na\r\n"u8;
+
+        // Tier 1b: Scalar ulong switch (variable-arg commands)
+        static ReadOnlySpan<byte> CMD_SETEX => "*4\r\n$5\r\nSETEX\r\n$1\r\na\r\n$2\r\n60\r\n$1\r\nb\r\n"u8;
+        static ReadOnlySpan<byte> CMD_EXPIRE => "*3\r\n$6\r\nEXPIRE\r\n$1\r\na\r\n$2\r\n60\r\n"u8;
+
+        // Old Tier 2 (FastParseArrayCommand): near top of switch chains (short names, common first chars)
+        static ReadOnlySpan<byte> CMD_HSET => "*4\r\n$4\r\nHSET\r\n$1\r\nh\r\n$1\r\nf\r\n$1\r\nv\r\n"u8;
+        static ReadOnlySpan<byte> CMD_LPUSH => "*3\r\n$5\r\nLPUSH\r\n$1\r\nl\r\n$1\r\nv\r\n"u8;
+        static ReadOnlySpan<byte> CMD_ZADD => "*4\r\n$4\r\nZADD\r\n$1\r\nz\r\n$1\r\n1\r\n$1\r\nm\r\n"u8;
+
+        // Old Tier 2 (FastParseArrayCommand): deep in switch chains (long names, double-digit $ header)
+        static ReadOnlySpan<byte> CMD_ZRANGEBYSCORE => "*4\r\n$13\r\nZRANGEBYSCORE\r\n$1\r\nz\r\n$1\r\n0\r\n$2\r\n10\r\n"u8;
+        static ReadOnlySpan<byte> CMD_ZREMRANGEBYSCORE => "*4\r\n$16\r\nZREMRANGEBYSCORE\r\n$1\r\nz\r\n$1\r\n0\r\n$2\r\n10\r\n"u8;
+        static ReadOnlySpan<byte> CMD_HINCRBYFLOAT => "*4\r\n$12\r\nHINCRBYFLOAT\r\n$1\r\nh\r\n$1\r\nf\r\n$3\r\n1.5\r\n"u8;
+
+        // Old Tier 3 (SlowParseCommand): sequential SequenceEqual scan
+        static ReadOnlySpan<byte> CMD_SUBSCRIBE => "*2\r\n$9\r\nSUBSCRIBE\r\n$2\r\nch\r\n"u8;
+        static ReadOnlySpan<byte> CMD_GEORADIUS => "*6\r\n$9\r\nGEORADIUS\r\n$1\r\ng\r\n$1\r\n0\r\n$1\r\n0\r\n$3\r\n100\r\n$2\r\nkm\r\n"u8;
+        static ReadOnlySpan<byte> CMD_SETIFMATCH => "*4\r\n$10\r\nSETIFMATCH\r\n$1\r\na\r\n$1\r\nb\r\n$1\r\n0\r\n"u8;
+
+        // Pre-allocated buffers (pinned for pointer stability)
+        byte[] bufPing, bufGet, bufSet, bufIncr, bufExists, bufSetex, bufExpire, bufHset, bufLpush, bufZadd, bufSubscribe;
+        byte[] bufZrangebyscore, bufZremrangebyscore, bufHincrbyfloat, bufGeoradius, bufSetifmatch;
+
+        public override void GlobalSetup()
+        {
+            base.GlobalSetup();
+
+            // Pre-seed a key so GET/EXISTS don't return NOT_FOUND
+            SlowConsumeMessage("*3\r\n$3\r\nSET\r\n$1\r\na\r\n$1\r\nb\r\n"u8);
+
+            bufPing = GC.AllocateArray<byte>(CMD_PING.Length, pinned: true);
+            CMD_PING.CopyTo(bufPing);
+            bufGet = GC.AllocateArray<byte>(CMD_GET.Length, pinned: true);
+            CMD_GET.CopyTo(bufGet);
+            bufSet = GC.AllocateArray<byte>(CMD_SET.Length, pinned: true);
+            CMD_SET.CopyTo(bufSet);
+            bufIncr = GC.AllocateArray<byte>(CMD_INCR.Length, pinned: true);
+            CMD_INCR.CopyTo(bufIncr);
+            bufExists = GC.AllocateArray<byte>(CMD_EXISTS.Length, pinned: true);
+            CMD_EXISTS.CopyTo(bufExists);
+            bufSetex = GC.AllocateArray<byte>(CMD_SETEX.Length, pinned: true);
+            CMD_SETEX.CopyTo(bufSetex);
+            bufExpire = GC.AllocateArray<byte>(CMD_EXPIRE.Length, pinned: true);
+            CMD_EXPIRE.CopyTo(bufExpire);
+            bufHset = GC.AllocateArray<byte>(CMD_HSET.Length, pinned: true);
+            CMD_HSET.CopyTo(bufHset);
+            bufLpush = GC.AllocateArray<byte>(CMD_LPUSH.Length, pinned: true);
+            CMD_LPUSH.CopyTo(bufLpush);
+            bufZadd = GC.AllocateArray<byte>(CMD_ZADD.Length, pinned: true);
+            CMD_ZADD.CopyTo(bufZadd);
+            bufSubscribe = GC.AllocateArray<byte>(CMD_SUBSCRIBE.Length, pinned: true);
+            CMD_SUBSCRIBE.CopyTo(bufSubscribe);
+            bufZrangebyscore = GC.AllocateArray<byte>(CMD_ZRANGEBYSCORE.Length, pinned: true);
+            CMD_ZRANGEBYSCORE.CopyTo(bufZrangebyscore);
+            bufZremrangebyscore = GC.AllocateArray<byte>(CMD_ZREMRANGEBYSCORE.Length, pinned: true);
+            CMD_ZREMRANGEBYSCORE.CopyTo(bufZremrangebyscore);
+            bufHincrbyfloat = GC.AllocateArray<byte>(CMD_HINCRBYFLOAT.Length, pinned: true);
+            CMD_HINCRBYFLOAT.CopyTo(bufHincrbyfloat);
+            bufGeoradius = GC.AllocateArray<byte>(CMD_GEORADIUS.Length, pinned: true);
+            CMD_GEORADIUS.CopyTo(bufGeoradius);
+            bufSetifmatch = GC.AllocateArray<byte>(CMD_SETIFMATCH.Length, pinned: true);
+            CMD_SETIFMATCH.CopyTo(bufSetifmatch);
+        }
+
+        // === Tier 1a: SIMD Vector128 fast path ===
+
+        [Benchmark]
+        public RespCommand ParsePING()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufPing);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseGET()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufGet);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseSET()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufSet);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseINCR()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufIncr);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseEXISTS()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufExists);
+            return result;
+        }
+
+        // === Tier 1b: Scalar ulong switch ===
+
+        [Benchmark]
+        public RespCommand ParseSETEX()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufSetex);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseEXPIRE()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufExpire);
+            return result;
+        }
+
+        // === Old Tier 2 (FastParseArrayCommand): near top of switch ===
+
+        [Benchmark]
+        public RespCommand ParseHSET()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufHset);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseLPUSH()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufLpush);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseZADD()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufZadd);
+            return result;
+        }
+
+        // === Old Tier 2 (FastParseArrayCommand): deep in switch (long names) ===
+
+        [Benchmark]
+        public RespCommand ParseZRANGEBYSCORE()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufZrangebyscore);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseZREMRANGEBYSCORE()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufZremrangebyscore);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseHINCRBYFLOAT()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufHincrbyfloat);
+            return result;
+        }
+
+        // === Old Tier 3 (SlowParseCommand): sequential SequenceEqual scan ===
+
+        [Benchmark]
+        public RespCommand ParseSUBSCRIBE()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufSubscribe);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseGEORADIUS()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufGeoradius);
+            return result;
+        }
+
+        [Benchmark]
+        public RespCommand ParseSETIFMATCH()
+        {
+            RespCommand result = default;
+            for (int i = 0; i < batchSize; i++)
+                result = session.ParseRespCommandBuffer(bufSetifmatch);
+            return result;
+        }
+    }
+}