diff --git a/README.md b/README.md
index e53cf28f5..2e6f74841 100644
--- a/README.md
+++ b/README.md
@@ -26,5 +26,3 @@ cd build-opt && ninja midi-redis
 ```
 
 for more options, run `./midi-redis --help`
-
-
diff --git a/helio b/helio
index 29e247c00..dab0fc768 160000
--- a/helio
+++ b/helio
@@ -1 +1 @@
-Subproject commit 29e247c00c2f10798c78044607436953fb22b384
+Subproject commit dab0fc768494d8e7cc42d77b08fbee1a81415072
diff --git a/server/dragonfly_connection.cc b/server/dragonfly_connection.cc
index 56a7467d9..72dbdffad 100644
--- a/server/dragonfly_connection.cc
+++ b/server/dragonfly_connection.cc
@@ -26,8 +26,6 @@ namespace fibers = boost::fibers;
 namespace dfly {
 namespace {
 
-using CmdArgVec = std::vector<MutableStrSpan>;
-
 void SendProtocolError(RedisParser::Result pres, FiberSocketBase* peer) {
   string res("-ERR Protocol error: ");
   if (pres == RedisParser::BAD_BULKLEN) {
@@ -43,10 +41,6 @@ void SendProtocolError(RedisParser::Result pres, FiberSocketBase* peer) {
   }
 }
 
-inline MutableStrSpan ToMSS(absl::Span<uint8_t> span) {
-  return MutableStrSpan{reinterpret_cast<char*>(span.data()), span.size()};
-}
-
 void RespToArgList(const RespVec& src, CmdArgVec* dest) {
   dest->resize(src.size());
   for (size_t i = 0; i < src.size(); ++i) {
@@ -311,6 +305,7 @@ void Connection::DispatchFiber(util::FiberSocketBase* peer) {
     std::unique_ptr<Request> req{dispatch_q_.front()};
     dispatch_q_.pop_front();
 
+    cc_->SetBatchMode(!dispatch_q_.empty());
     cc_->conn_state.mask |= ConnectionState::ASYNC_DISPATCH;
     service_->DispatchCommand(CmdArgList{req->args.data(), req->args.size()}, cc_.get());
     cc_->conn_state.mask &= ~ConnectionState::ASYNC_DISPATCH;
diff --git a/server/dragonfly_listener.cc b/server/dragonfly_listener.cc
index d5fa73672..dd1e80b2a 100644
--- a/server/dragonfly_listener.cc
+++ b/server/dragonfly_listener.cc
@@ -119,11 +119,12 @@ ProactorBase* Listener::PickConnectionProactor(LinuxSocketBase* sock) {
   if (FLAGS_conn_use_incoming_cpu) {
     int fd = sock->native_handle();
 
-    int cpu;
+    int cpu, napi_id;
     socklen_t len = sizeof(cpu);
 
     CHECK_EQ(0, getsockopt(fd, SOL_SOCKET, SO_INCOMING_CPU, &cpu, &len));
-    VLOG(1) << "CPU for connection " << fd << " is " << cpu;
+    CHECK_EQ(0, getsockopt(fd, SOL_SOCKET, SO_INCOMING_NAPI_ID, &napi_id, &len));
+    VLOG(1) << "CPU/NAPI for connection " << fd << " is " << cpu << "/" << napi_id;
 
     vector<unsigned> ids = pool()->MapCpuToThreads(cpu);
     if (!ids.empty()) {
diff --git a/server/engine_shard_set.cc b/server/engine_shard_set.cc
index 501cb5bfc..b8add1194 100644
--- a/server/engine_shard_set.cc
+++ b/server/engine_shard_set.cc
@@ -43,7 +43,7 @@ void EngineShard::DestroyThreadLocal() {
   delete shard_;
   shard_ = nullptr;
 
-  DVLOG(1) << "Shard reset " << index;
+  VLOG(1) << "Shard reset " << index;
 }
 
 void EngineShardSet::Init(uint32_t sz) {
diff --git a/server/main_service.cc b/server/main_service.cc
index a54943b4d..f7ba55c1a 100644
--- a/server/main_service.cc
+++ b/server/main_service.cc
@@ -65,6 +65,8 @@ void Service::Init(util::AcceptServer* acceptor) {
 }
 
 void Service::Shutdown() {
+  VLOG(1) << "Service::Shutdown";
+
   engine_varz.reset();
   request_latency_usec.Shutdown();
   ping_qps.Shutdown();
diff --git a/server/reply_builder.cc b/server/reply_builder.cc
index 68af563d7..c4e56862c 100644
--- a/server/reply_builder.cc
+++ b/server/reply_builder.cc
@@ -30,7 +30,30 @@ BaseSerializer::BaseSerializer(io::Sink* sink) : sink_(sink) {
 }
 
 void BaseSerializer::Send(const iovec* v, uint32_t len) {
-  error_code ec = sink_->Write(v, len);
+  if (should_batch_) {
+    // TODO: to introduce flushing when too much data is batched.
+    for (unsigned i = 0; i < len; ++i) {
+      std::string_view src((char*)v[i].iov_base, v[i].iov_len);
+      DVLOG(2) << "Appending to stream " << sink_ << " " << src;
+      batch_.append(src.data(), src.size());
+    }
+    return;
+  }
+
+  error_code ec;
+  if (batch_.empty()) {
+    ec = sink_->Write(v, len);
+  } else {
+    DVLOG(1) << "Sending batch to stream " << sink_ << "\n" << batch_;
+
+    iovec tmp[len + 1];
+    tmp[0].iov_base = batch_.data();
+    tmp[0].iov_len = batch_.size();
+    copy(v, v + len, tmp + 1);
+    ec = sink_->Write(tmp, len + 1);
+    batch_.clear();
+  }
+
   if (ec) {
     ec_ = ec;
   }
diff --git a/server/reply_builder.h b/server/reply_builder.h
index 667f8b835..06837fe62 100644
--- a/server/reply_builder.h
+++ b/server/reply_builder.h
@@ -22,18 +22,24 @@ class BaseSerializer {
       ec_ = std::make_error_code(std::errc::connection_aborted);
   }
 
+  // In order to reduce interrupt rate we allow coalescing responses together using
+  // Batch mode. It is controlled by Connection state machine because it makes sense only
+  // when pipelined requests are arriving.
+  void SetBatchMode(bool batch) {
+    should_batch_ = batch;
+  }
+
   //! Sends a string as is without any formatting. raw should be encoded according to the protocol.
   void SendDirect(std::string_view str);
 
-  ::io::Sink* sink() {
-    return sink_;
-  }
-
   void Send(const iovec* v, uint32_t len);
 
  private:
   ::io::Sink* sink_;
   std::error_code ec_;
+  std::string batch_;
+
+  bool should_batch_ = false;
 };
 
 class RespSerializer : public BaseSerializer {
@@ -89,6 +95,10 @@ class ReplyBuilder {
   void SendGetReply(std::string_view key, uint32_t flags, std::string_view value);
   void SendGetNotFound();
 
+  void SetBatchMode(bool mode) {
+    serializer_->SetBatchMode(mode);
+  }
+
  private:
   RespSerializer* as_resp() {
     return static_cast<RespSerializer*>(serializer_.get());