diff --git a/.gitignore b/.gitignore
index 0fd2fc2dbe9..4aadccbc9d3 100644
--- a/.gitignore
+++ b/.gitignore
@@ -32,6 +32,7 @@ target/
 /agent/src/ebpf/kernel/*.ll
 /agent/src/ebpf/kernel/*.objdump
 /agent/src/ebpf/kernel/*.elf
+/agent/src/ebpf/kernel/build/
 /agent/src/ebpf/user/socket_trace_bpf_5_2.c
 /agent/src/ebpf/user/socket_trace_bpf_core.c
 /agent/src/ebpf/user/socket_trace_bpf_common.c
@@ -39,6 +40,7 @@ target/
 /agent/src/ebpf/user/socket_trace_bpf_kylin.c
 /agent/src/ebpf/user/perf_profiler_bpf_common.c
 /agent/src/ebpf/user/af_packet_fanout_bpf_common.c
+/agent/src/ebpf/user/extended/*_bpf_common.c
 /agent/src/ebpf/user/deepflow_ebpfctl_bin.c
 /agent/src/ebpf/user/profile/java_agent_so_gnu.c
 /agent/src/ebpf/user/profile/java_agent_so_musl.c
diff --git a/agent/src/config/config.rs b/agent/src/config/config.rs
index 7c36a017a7a..3940b09e98a 100644
--- a/agent/src/config/config.rs
+++ b/agent/src/config/config.rs
@@ -23,8 +23,15 @@ use std::net::{IpAddr, ToSocketAddrs};
 use std::path::Path;
 use std::time::Duration;
 
+#[cfg(feature = "extended_observability")]
+use std::ffi::CString;
+
 #[cfg(any(target_os = "linux", target_os = "android"))]
 use envmnt::{ExpandOptions, ExpansionType};
+#[cfg(feature = "extended_observability")]
+use libc::c_int;
+#[cfg(feature = "extended_observability")]
+use log::warn;
 use log::{debug, error, info};
 use md5::{Digest, Md5};
 #[cfg(any(target_os = "linux", target_os = "android"))]
@@ -39,6 +46,9 @@ use tokio::runtime::Runtime;
 
 use crate::common::l7_protocol_log::{L7ProtocolBitmap, L7ProtocolParser};
 use crate::dispatcher::recv_engine::DEFAULT_BLOCK_SIZE;
+#[cfg(any(target_os = "linux", target_os = "android"))]
+#[cfg(feature = "extended_observability")]
+use crate::ebpf;
 use crate::flow_generator::{DnsLog, MemcachedLog};
 #[cfg(any(target_os = "linux", target_os = "android"))]
 use crate::platform::{OsAppTag, ProcessData};
@@ -1235,6 +1245,97 @@ impl Default for EbpfTunning {
     }
 }
 
+#[derive(Clone, Debug, Deserialize, PartialEq, Eq)]
+#[serde(default)]
+pub struct NicOptimizeConfig {
+    pub interface: String,
+    pub rx_ring_size: u64,
+    pub rss_channel_count: u64,
+    pub irq_cpu_list: String,
+    pub xdp_cpu_redirect: bool,
+    pub xdp_queue_size: u64,
+    pub xdp_cpu_redirect_list: String,
+}
+
+const XDP_QUEUE_SIZE_MIN: u64 = 512;
+const XDP_QUEUE_SIZE_MAX: u64 = 8192;
+
+impl Default for NicOptimizeConfig {
+    fn default() -> Self {
+        Self {
+            interface: "".to_string(),
+            rx_ring_size: 0,
+            rss_channel_count: 0,
+            irq_cpu_list: "".to_string(),
+            xdp_cpu_redirect: false,
+            xdp_queue_size: 2048,
+            xdp_cpu_redirect_list: "".to_string(),
+        }
+    }
+}
+
+impl NicOptimizeConfig {
+    #[cfg(any(target_os = "linux", target_os = "android"))]
+    #[cfg(feature = "extended_observability")]
+    pub fn apply(&self) {
+        let to_cstring = |field: &str, value: &str| {
+            CString::new(value)
+                .map_err(|_| {
+                    warn!(
+                        "Skip NIC optimization for interface {:?}: {} contains NUL byte",
+                        self.interface, field
+                    );
+                })
+                .ok()
+        };
+
+        let Some(nic_name) = to_cstring("interface", self.interface.as_str()) else {
+            return;
+        };
+        let Some(irq_cpu) = to_cstring("irq_cpu_list", self.irq_cpu_list.as_str()) else {
+            return;
+        };
+        let Some(xdp_cpu) =
+            to_cstring("xdp_cpu_redirect_list", self.xdp_cpu_redirect_list.as_str())
+        else {
+            return;
+        };
+
+        let ret = unsafe {
+            ebpf::nic_optimize_config(
+                nic_name.as_ptr(),
+                self.rx_ring_size as c_int,
+                self.rss_channel_count as c_int,
+                irq_cpu.as_ptr(),
+                self.xdp_cpu_redirect,
+                self.xdp_queue_size as c_int,
+                xdp_cpu.as_ptr(),
+            )
+        };
+
+        if ret != 0 {
+            warn!(
+                "Failed to configure NIC optimization for interface '{}' \
+                (ret={}, rx_ring_size={}, rss_channel_count={}, \
+                xdp_cpu_redirect={}, xdp_queue_size={})",
+                self.interface,
+                ret,
+                self.rx_ring_size,
+                self.rss_channel_count,
+                self.xdp_cpu_redirect,
+                self.xdp_queue_size,
+            );
+        }
+    }
+}
+
+#[derive(Clone, Debug, Default, Deserialize, PartialEq, Eq)]
+#[serde(default)]
+pub struct EbpfNetwork {
+    pub nic_opt_enabled: bool,
+    pub nic_optimize: Vec<NicOptimizeConfig>,
+}
+
 #[derive(Clone, Debug, Deserialize, PartialEq, Eq)]
 #[serde(default)]
 pub struct EbpfSocketPreprocess {
@@ -1285,6 +1386,7 @@ pub struct Ebpf {
     pub file: EbpfFile,
     pub profile: EbpfProfile,
     pub tunning: EbpfTunning,
+    pub network: EbpfNetwork,
     #[serde(skip)]
     pub java_symbol_file_refresh_defer_interval: i32,
 }
@@ -1297,6 +1399,7 @@ impl Default for Ebpf {
             file: EbpfFile::default(),
             profile: EbpfProfile::default(),
             tunning: EbpfTunning::default(),
+            network: EbpfNetwork::default(),
             java_symbol_file_refresh_defer_interval: 60,
         }
     }
@@ -3219,6 +3322,15 @@ impl UserConfig {
             )));
         }
 
+        for nic in &self.inputs.ebpf.network.nic_optimize {
+            if !(XDP_QUEUE_SIZE_MIN..=XDP_QUEUE_SIZE_MAX).contains(&nic.xdp_queue_size) {
+                return Err(ConfigError::RuntimeConfigInvalid(format!(
+                    "xdp_queue_size {} for interface {} not in [{}, {}]",
+                    nic.xdp_queue_size, nic.interface, XDP_QUEUE_SIZE_MIN, XDP_QUEUE_SIZE_MAX,
+                )));
+            }
+        }
+
         Ok(())
     }
 
diff --git a/agent/src/config/handler.rs b/agent/src/config/handler.rs
index 2b6468c3bfd..dc361cd8d53 100755
--- a/agent/src/config/handler.rs
+++ b/agent/src/config/handler.rs
@@ -3553,6 +3553,26 @@ impl ConfigHandler {
             restart_agent = !first_run;
         }
 
+        // The hot-update configuration items in ebpf.network are handled within the set_ebpf() callback.
+        let network = &mut ebpf.network;
+        let new_network = &mut new_ebpf.network;
+
+        if network.nic_opt_enabled != new_network.nic_opt_enabled {
+            info!(
+                "Update inputs.ebpf.network.nic_opt_enabled from {:?} to {:?}.",
+                network.nic_opt_enabled, new_network.nic_opt_enabled
+            );
+            network.nic_opt_enabled = new_network.nic_opt_enabled;
+        }
+
+        if network.nic_optimize != new_network.nic_optimize {
+            info!(
+                "Update inputs.ebpf.network.nic_optimize from {:?} to {:?}.",
+                network.nic_optimize, new_network.nic_optimize
+            );
+            network.nic_optimize = new_network.nic_optimize.clone();
+        }
+
         let kprobe = &mut ebpf.socket.kprobe;
         let new_kprobe = &mut new_ebpf.socket.kprobe;
         if kprobe.disabled != new_kprobe.disabled {
diff --git a/agent/src/ebpf/Makefile b/agent/src/ebpf/Makefile
index 687c5a7bb45..10d4421f710 100644
--- a/agent/src/ebpf/Makefile
+++ b/agent/src/ebpf/Makefile
@@ -158,7 +158,7 @@ PERF_PROFILER_ELFS := user/perf_profiler_bpf_common.c \
 	user/perf_profiler_bpf_5_2_plus.c \
 	user/perf_profiler_bpf_5_15_plus.c
 
-ELFFILES := $(SOCKET_TRACE_ELFS) $(PERF_PROFILER_ELFS) $(AF_PACKET_FANOUT_ELFS) $(TCP_OPTION_TRACING_ELFS)
+ELFFILES := $(SOCKET_TRACE_ELFS) $(PERF_PROFILER_ELFS) $(EXTEND_ELFS)
 
 tools/bintobuffer:
 	$(call msg,TOOLS,tools/bintobuffer)
diff --git a/agent/src/ebpf/kernel/Makefile b/agent/src/ebpf/kernel/Makefile
index 5173006740f..8f3318419d5 100644
--- a/agent/src/ebpf/kernel/Makefile
+++ b/agent/src/ebpf/kernel/Makefile
@@ -60,7 +60,7 @@ FINAL_TARGET = -emit-llvm -D__TARGET_ARCH_$(ARCH) -o ${@:.elf=.ll} -c $^ && $(LL
 
 all: $(TAEGET_KERN_ELF)
 
-EBPF_CLAGS ?= -I. -Ivmlinux -Iinclude -I../../../crates/trace-utils/src \
+EBPF_CLAGS ?= -I. -Ivmlinux -Iinclude -I../user/extended/bpf/include -I../../../crates/trace-utils/src \
         -D__BPF_TRACING__ -D GROUP_LEADER_OFFSET_OVERRIDE=0 \
         -DSTART_BOOTTIME_OFFSET_OVERRIDE=0      \
         -DSTART_BOOTTIME_VARNAME=real_start_time
diff --git a/agent/src/ebpf/kernel/include/bpf_base.h b/agent/src/ebpf/kernel/include/bpf_base.h
index d13e812121d..d2f3d05a617 100644
--- a/agent/src/ebpf/kernel/include/bpf_base.h
+++ b/agent/src/ebpf/kernel/include/bpf_base.h
@@ -30,7 +30,6 @@
 #include <errno.h>
 #include <stddef.h>
 #include <bcc/compat/linux/bpf.h>
-#include "utils.h"
 
 struct task_struct;
 
@@ -150,6 +149,13 @@ static int
     __attribute__ ((__unused__)) (*bpf_get_stack) (void *ctx, void *buf, __u32 size,
 						     int flags) = (void *)67;
 
+
+// Linux 4.14: Added support for BPF_MAP_TYPE_CPUMAP, allowing packets to be redirected to specific CPUs.
+static long
+    __attribute__ ((__unused__)) (*bpf_redirect_map)(void *map, __u32 key, __u64 flags) =
+							(void *)51;
+
+
 /* llvm builtin functions that eBPF C program may use to 
  * emit BPF_LD_ABS and BPF_LD_IND instructions 
  */
@@ -334,6 +340,7 @@ _Pragma("GCC error \"PT_GO_REGS_PARM\"");
 
 #define TP_SYSCALL_PROG(F) SEC("tracepoint/syscalls/sys_"__stringify(F)) int df_T_##F
 #define TP_SCHED_PROG(F) SEC("tracepoint/sched/sched_"__stringify(F)) int df_T_##F
+#define TP_XDP_PROG(F) SEC("tracepoint/xdp/xdp_"__stringify(F)) int df_T_##F
 #define PROGTP(F) SEC("prog/tp/"__stringify(F)) int df_TP_##F
 #define PROGKP(F) SEC("prog/kp/"__stringify(F)) int df_KP_##F
 #define PROGPE(F) SEC("prog/pe/"__stringify(F)) int df_PE_##F
diff --git a/agent/src/ebpf/kernel/perf_profiler.bpf.c b/agent/src/ebpf/kernel/perf_profiler.bpf.c
index f1c7f870f65..3b743e23d7e 100644
--- a/agent/src/ebpf/kernel/perf_profiler.bpf.c
+++ b/agent/src/ebpf/kernel/perf_profiler.bpf.c
@@ -19,6 +19,7 @@
  * SPDX-License-Identifier: GPL-2.0
  */
 
+#include <arpa/inet.h>
 #include <linux/bpf_perf_event.h>
 #include "config.h"
 #include "bpf_base.h"
diff --git a/agent/src/ebpf/mod.rs b/agent/src/ebpf/mod.rs
index e47bb80995c..e2e9a37cadf 100644
--- a/agent/src/ebpf/mod.rs
+++ b/agent/src/ebpf/mod.rs
@@ -941,6 +941,57 @@ extern "C" {
                 */
                 pub fn set_tcp_option_tracing_enabled(enabled: bool) -> c_int;
                 pub fn set_tcp_option_tracing_sample_window(bytes: c_uint) -> c_int;
+
+                /**
+                 * @brief Enable or disable NIC optimization.
+                 *
+                 * This function sets the NIC optimization feature on or off. When enabled,
+                 * the system may adjust RX ring sizes, RSS channels, and CPU affinities
+                 * to improve packet processing performance.
+                 *
+                 * @param enabled Set to 1 to enable, or 0 to disable. On the Rust side,
+                 *                this can be converted from a bool.
+                 *
+                 * @return 0 on success, non-zero on failure.
+                 *
+                 * @note Corresponding C function: `int set_nic_optimization(bool enabled)`
+                 */
+                pub fn set_nic_optimization(enabled: bool) -> c_int;
+
+                /**
+                 * @brief Configure optimization parameters for a single NIC.
+                 *
+                 * This function applies NIC-specific optimization settings, including
+                 * RX ring size, RSS channel count, IRQ CPU binding, and XDP CPU redirection.
+                 * It allows fine-grained performance tuning for each network interface.
+                 *
+                 * @param nic_name The name of the network interface (C string).
+                 * @param rx_ring_size The size of the RX ring buffer.
+                 * @param rss_channel_count Number of RSS channels to use.
+                 * @param irq_cpu_list Comma-separated list of CPU IDs for IRQ binding.
+                 * @param xdp_cpu_redirect Set to 1 to enable XDP CPU redirect, 0 to disable.
+                 * @param xdp_queue_size The size of the XDP queue.
+                 * @param xdp_cpu_list Comma-separated list of CPU IDs for XDP processing.
+                 *
+                 * @return 0 on success, non-zero on failure.
+                 *
+                 * @note Corresponding C function:
+                 * ```c
+                 * int nic_optimize_config(const char* nic_name, int rx_ring_size,
+                 *                         int rss_channel_count, const char* irq_cpu_list,
+                 *                         bool xdp_cpu_redirect, int xdp_queue_size,
+                 *                         const char* xdp_cpu_list);
+                 * ```
+                 */
+                pub fn nic_optimize_config(
+                    nic_name: *const c_char,
+                    rx_ring_size: c_int,
+                    rss_channel_count: c_int,
+                    irq_cpu_list: *const c_char,
+                    xdp_cpu_redirect: bool,
+                    xdp_queue_size: c_int,
+                    xdp_cpu_list: *const c_char,
+                ) -> c_int;
         }
     }
 }
diff --git a/agent/src/ebpf/samples/rust/socket-tracer/Cargo.toml b/agent/src/ebpf/samples/rust/socket-tracer/Cargo.toml
index 0b691165def..2e7fd8f07cd 100644
--- a/agent/src/ebpf/samples/rust/socket-tracer/Cargo.toml
+++ b/agent/src/ebpf/samples/rust/socket-tracer/Cargo.toml
@@ -21,3 +21,5 @@ lazy_static = "1.4.0"
 log = "0.4"
 env_logger = "0.9"
 trace-utils = { path = "../../../../../crates/trace-utils" }
+# Enable this option when building the Enterprise edition.
+# trace-utils-interp = { path = "../../../../../crates/trace-utils-interp" }
diff --git a/agent/src/ebpf/samples/rust/socket-tracer/src/main.rs b/agent/src/ebpf/samples/rust/socket-tracer/src/main.rs
index 5eb4d1bc905..7361a87a892 100644
--- a/agent/src/ebpf/samples/rust/socket-tracer/src/main.rs
+++ b/agent/src/ebpf/samples/rust/socket-tracer/src/main.rs
@@ -28,6 +28,21 @@ use std::thread;
 use std::time::{Duration, UNIX_EPOCH};
 use log::info;
 
+// Reference trace-utils-interp when building the Enterprise edition.
+// The purpose is to ensure that Rust links against libtrace_utils_interp-xxxx.rlib
+// during the linking stage.
+//
+// interpreter_tracer.c calls is_php_process(),
+// and this function is defined in the trace-utils-interp crate.
+//
+// However, the Rust code in socket_tracer does not directly reference
+// trace-utils-interp. As a result, Cargo considers this dependency unused
+// and excludes it from the linking stage.
+//
+// Therefore, we explicitly reference trace-utils-interp here to force
+// Cargo to include it in the final link.
+//use trace_utils_interp as _;
+
 extern "C" {
     fn print_uprobe_http2_info(data: *mut c_char, len: c_uint);
     fn print_uprobe_grpc_dataframe(data: *mut c_char, len: c_uint);
@@ -746,6 +761,31 @@ fn main() {
             std::thread::sleep(Duration::from_secs(1));
         }
 
+        // ------ Nic Optimization ----
+        //if set_nic_optimization(true) != 0 {
+        //    println!("set_nic_optimization() error.");
+        //    ::std::process::exit(1);
+        //}
+
+        //let nic_name = CString::new("p2p2").unwrap();
+        //let irq_cpu_list = CString::new("").unwrap();
+        //let xdp_cpu_list = CString::new("").unwrap();
+
+        //if nic_optimize_config(
+        //    nic_name.as_c_str().as_ptr(),
+        //    0, // rx_ring_size
+        //    0, // rss_channel_count
+        //    irq_cpu_list.as_c_str().as_ptr(),
+        //    true, // xdp_cpu_redirect
+        //    0,    // xdp_queue_size
+        //    xdp_cpu_list.as_c_str().as_ptr(),
+        //) != 0
+        //{
+        //    println!("nic_optimize_config() error.");
+        //    ::std::process::exit(1);
+        //}
+        // ------ Nic Optimization end ----
+
         //thread::sleep(Duration::from_secs(60));
         //stop_continuous_profiler();
         //print!(
diff --git a/agent/src/ebpf/user/config.h b/agent/src/ebpf/user/config.h
index d99487e987f..6fdf59318d3 100644
--- a/agent/src/ebpf/user/config.h
+++ b/agent/src/ebpf/user/config.h
@@ -180,6 +180,7 @@ enum cfg_feature_idx {
 	FEATURE_PROFILE_PYTHON,
 	FEATURE_PROFILE_PHP,
 	FEATURE_PROFILE_V8,
+	FEATURE_CPU_BALANCER,
 	FEATURE_MAX,
 };
 
@@ -194,6 +195,7 @@ enum cfg_feature_idx {
 #define FEATURE_FLAG_PROFILE_PYTHON		(1 << FEATURE_PROFILE_PYTHON)
 #define FEATURE_FLAG_PROFILE_PHP		(1 << FEATURE_PROFILE_PHP)
 #define FEATURE_FLAG_PROFILE_V8			(1 << FEATURE_PROFILE_V8)
+#define FEATURE_FLAG_CPU_BALANCER		(1 << FEATURE_CPU_BALANCER)
 
 #define FEATURE_FLAG_PROFILE				(FEATURE_FLAG_PROFILE_ONCPU | FEATURE_FLAG_PROFILE_OFFCPU | FEATURE_FLAG_PROFILE_MEMORY)
 
diff --git a/agent/src/ebpf/user/load.c b/agent/src/ebpf/user/load.c
index 549e31eb971..2600b209b54 100644
--- a/agent/src/ebpf/user/load.c
+++ b/agent/src/ebpf/user/load.c
@@ -630,6 +630,8 @@ static enum bpf_prog_type get_prog_type(struct sec_desc *desc)
 		prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
 	} else if (!strncmp(desc->name, "sockops", 7)) {
 		prog_type = BPF_PROG_TYPE_SOCK_OPS;
+	} else if (!memcmp(desc->name, "xdp", 3)) {
+		prog_type = BPF_PROG_TYPE_XDP;
 	} else {
 		prog_type = BPF_PROG_TYPE_UNSPEC;
 	}
diff --git a/agent/src/ebpf/user/probe.c b/agent/src/ebpf/user/probe.c
index 2f48734d32b..3c66d8e7964 100644
--- a/agent/src/ebpf/user/probe.c
+++ b/agent/src/ebpf/user/probe.c
@@ -73,6 +73,8 @@ int bpf_get_program_fd(void *obj, const char *name, void **p)
 			__name += 4;
 		} else if (strncmp(__name, "sched_", 6) == 0) {
 			__name += 6;
+		} else if (strncmp(__name, "xdp_", 4) == 0) {
+			__name += 4;
 		}
 
 		res =
diff --git a/agent/src/ebpf/user/tracer.c b/agent/src/ebpf/user/tracer.c
index 798a20f5cce..cfce1a79800 100644
--- a/agent/src/ebpf/user/tracer.c
+++ b/agent/src/ebpf/user/tracer.c
@@ -140,7 +140,6 @@ pthread_mutex_t match_pids_lock;
 // Used to store process IDs for matching various features
 pids_match_hash_t pids_match_hash;
 
-static int tracepoint_attach(struct tracepoint *tp);
 static int perf_reader_setup(struct bpf_perf_reader *perf_readerm,
 			     int thread_nr);
 static void perf_reader_release(struct bpf_perf_reader *perf_reader);
@@ -1020,7 +1019,7 @@ int probe_detach(struct probe *p)
 	return ret;
 }
 
-static int tracepoint_attach(struct tracepoint *tp)
+int tracepoint_attach(struct tracepoint *tp)
 {
 	if (tp->link) {
 		return ETR_EXIST;
@@ -1039,7 +1038,7 @@ static int tracepoint_attach(struct tracepoint *tp)
 	return ETR_OK;
 }
 
-static int tracepoint_detach(struct tracepoint *tp)
+int tracepoint_detach(struct tracepoint *tp)
 {
 	if (tp->link == NULL) {
 		return ETR_NOTEXIST;
diff --git a/agent/src/ebpf/user/tracer.h b/agent/src/ebpf/user/tracer.h
index fc6634d254d..aa2e53e6c79 100644
--- a/agent/src/ebpf/user/tracer.h
+++ b/agent/src/ebpf/user/tracer.h
@@ -755,4 +755,41 @@ struct ebpf_object *create_ebpf_object(const void *bpf_code,
  * @return 0 on success, or a non-zero value on error.
  */
 int load_ebpf_object(struct ebpf_object *obj);
+
+/**
+ * tracepoint_attach - Attach an eBPF program to a tracepoint
+ * @tp: Pointer to tracepoint descriptor
+ *
+ * This function attaches the eBPF program referenced by @tp->prog
+ * to its corresponding tracepoint and stores the returned link
+ * object into @tp->link.
+ *
+ * Return:
+ *   ETR_OK        - Attach success
+ *   ETR_EXIST     - Tracepoint already attached
+ *   ETR_INVAL     - Attach failed
+ *
+ * Notes:
+ *   - This function is NOT fully thread-safe. Caller must ensure
+ *     proper synchronization if called concurrently.
+ *   - On failure, attach_failed_count will be incremented.
+ */
+int tracepoint_attach(struct tracepoint *tp);
+
+/**
+ * tracepoint_detach - Detach an eBPF program from a tracepoint
+ * @tp: Pointer to tracepoint descriptor
+ *
+ * This function detaches the eBPF program associated with the
+ * given tracepoint and releases the link object.
+ *
+ * Return:
+ *   ETR_OK         - Detach success
+ *   ETR_NOTEXIST   - Tracepoint not attached
+ *
+ * Notes:
+ *   - Safe to call even if detach callback is NULL.
+ *   - Caller must ensure no concurrent attach/detach.
+ */
+int tracepoint_detach(struct tracepoint *tp);
 #endif /* DF_USER_TRACER_H */
diff --git a/agent/src/ebpf/user/utils.h b/agent/src/ebpf/user/utils.h
index 1d53ef6b9e7..7188120f745 100644
--- a/agent/src/ebpf/user/utils.h
+++ b/agent/src/ebpf/user/utils.h
@@ -48,6 +48,8 @@
 
 #define MAX_PATH_LENGTH 1024
 #define CONTAINER_ID_SIZE 65
+#ifndef _LINUX_SYSINFO_H   /* Header guard used by glibc and musl */
+#define _LINUX_SYSINFO_H
 
 struct sysinfo {
 	long uptime;
@@ -66,6 +68,8 @@ struct sysinfo {
 	char _f[20 - 2 * sizeof(unsigned long) - sizeof(uint32_t)];
 };
 
+#endif
+
 extern int sysinfo(struct sysinfo *__info);
 
 #ifndef likely
diff --git a/agent/src/ebpf_dispatcher.rs b/agent/src/ebpf_dispatcher.rs
index 8c471fd1c84..f5d61997f50 100644
--- a/agent/src/ebpf_dispatcher.rs
+++ b/agent/src/ebpf_dispatcher.rs
@@ -1049,6 +1049,18 @@ impl EbpfCollector {
                     tcp_option_trace.enabled
                 );
             }
+
+            // NicOptimize
+            if ebpf::set_nic_optimization(config.ebpf.network.nic_opt_enabled) != 0 {
+                warn!(
+                    "Failed to apply NIC optimization setting (enabled: {})",
+                    config.ebpf.network.nic_opt_enabled
+                );
+            }
+
+            for nic in &config.ebpf.network.nic_optimize {
+                nic.apply();
+            }
         }
 
         if ebpf::running_socket_tracer(
@@ -1526,8 +1538,21 @@ impl EbpfCollector {
                         tcp_option_trace.enabled
                     );
                 }
+
+                // NicOptimize
+                if ebpf::set_nic_optimization(config.ebpf.network.nic_opt_enabled) != 0 {
+                    warn!(
+                        "Failed to apply NIC optimization setting (enabled: {})",
+                        config.ebpf.network.nic_opt_enabled
+                    );
+                }
+
+                for nic in &config.ebpf.network.nic_optimize {
+                    nic.apply();
+                }
             }
         }
+
         if config.l7_log_enabled() || config.dpdk_enabled {
             self.start();
         } else {
diff --git a/server/agent_config/README-CH.md b/server/agent_config/README-CH.md
index 1218f86bf16..6e57e7574c6 100644
--- a/server/agent_config/README-CH.md
+++ b/server/agent_config/README-CH.md
@@ -5165,6 +5165,362 @@ inputs:
 禁用 Node.js（V8）解释器剖析。禁用后将不采集 Node.js 进程的函数调用栈，
 可节省约 6.4 MB 内核内存（v8_unwind_info_map）。
 
+### 网络 {#inputs.ebpf.network}
+
+#### NIC optimization Enabled {#inputs.ebpf.network.nic_opt_enabled}
+
+**标签**:
+
+`hot_update`
+<mark>ee_feature</mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_opt_enabled`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_opt_enabled: false
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | bool |
+
+**详细描述**:
+
+是否启用网卡优化功能，用于提升多核环境下的网络包处理能力
+以及突发流量承载能力。
+
+开启后将综合进行以下优化：
+  - RSS 硬件队列数量配置
+  - RX ring 描述符数量调优
+  - 硬件中断与 CPU 亲和性绑定（IRQ 绑核）
+  - 可选的 XDP CPUMAP 软件重定向分发
+
+该优化主要解决 RSS 硬件无法基于封装报文内层头部
+（如 GRE、Double VLAN、VXLAN、ERSPAN）进行哈希分摊的问题，
+避免流量集中在单个 CPU 上造成过载和丢包。
+
+通过调整 RX ring 描述符数量，可提升突发流量场景下的
+接收缓存能力，降低 ring 满导致的丢包风险。
+
+在启用 XDP CPU Redirect 时，数据包会在接收后通过
+CPUMAP 在多个 CPU 核心之间进行软件层重分发，
+从而突破硬件 RSS 的能力限制，实现更均衡的负载分布。
+
+建议在以下场景开启该功能：
+  1）使用 tcpdump 抓包发现该接口流量主要为 GRE、
+     Double VLAN、VXLAN 等封装报文；
+  2）通过 `top`（按 1）观察各 CPU 软中断占用率时，
+     发现某一个 CPU 的 softirq 接近 100%，
+     而其他 CPU 软中断占用率明显偏低。
+
+为获得最佳性能，建议将中断 CPU 和 XDP 重定向 CPU
+配置在与物理网卡相同的 NUMA 节点上。
+
+#### 网卡优化配置 {#inputs.ebpf.network.nic_optimize}
+
+**标签**:
+
+`hot_update`
+<mark>ee_feature</mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - interface: ''
+        irq_cpu_list: ''
+        rss_channel_count: 0
+        rx_ring_size: 0
+        xdp_cpu_redirect: false
+        xdp_cpu_redirect_list: ''
+        xdp_queue_size: 2048
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | dict |
+
+**详细描述**:
+
+针对指定网卡接口进行性能优化配置。
+
+该功能通过调优 RSS 队列、中断绑核、RX ring 大小、
+以及可选的 XDP CPUMAP 重定向机制，提升多核扩展能力
+和突发流量承载能力。
+
+建议在以下场景开启：
+  - 接口流量主要为 GRE、Double VLAN、VXLAN、ERSPAN 等封装报文；
+  - 某个 CPU 的 softirq 占用率接近 100%，而其他 CPU 空闲。
+
+为了获得更好的性能，程序会自动停用 irqbalance 服务，以防止网卡中断在 CPU 之间迁移。
+
+可为多个接口分别配置优化策略。
+
+样例:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_opt_enabled: true
+      nic_optimize:
+      - interface: eth0
+        rx_ring_size: 4096
+        rss_channel_count: 2
+        irq_cpu_list: 1,2
+        xdp_cpu_redirect: true
+        xdp_queue_size: 2048
+        xdp_cpu_redirect_list: 4,5,6,7
+      - interface: eth1
+        rx_ring_size: 4096
+        rss_channel_count: 2
+        irq_cpu_list: 1,2
+        xdp_cpu_redirect: true
+        xdp_queue_size: 2048
+        xdp_cpu_redirect_list: 4,5,6,7
+```
+
+##### 网卡接口 {#inputs.ebpf.network.nic_optimize.interface}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.interface`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - interface: ''
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | string |
+
+**详细描述**:
+
+需要进行优化的网卡接口名称。
+
+##### RX Ring 描述符数量 {#inputs.ebpf.network.nic_optimize.rx_ring_size}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.rx_ring_size`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - rx_ring_size: 0
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | int |
+
+**详细描述**:
+
+网卡接收环（RX ring）的描述符数量。
+
+增大该值可提升突发流量场景下的缓存能力，
+降低因 ring 满导致的丢包风险。
+具体使用`ethtool -g <iface>` 查看当前配置，根据实际情况调整到合适的值。
+
+默认值为 0 表示保持原状忽略此项配置。
+
+##### RSS 队列数量 {#inputs.ebpf.network.nic_optimize.rss_channel_count}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.rss_channel_count`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - rss_channel_count: 0
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | int |
+
+**详细描述**:
+
+RSS 硬件队列数量。
+数据包在物理网卡完成哈希后，将分发到指定数量的队列并触发中断。
+
+最大一般支持 16，且不要超过逻辑 CPU 核数。
+具体使用`ethtool -l <iface>` 查看当前配置，根据实际情况调整到合适的值。
+
+当启用 XDP CPU Redirect 时建议设置为 1。
+默认值为 0 表示保持原状忽略此项配置。
+
+##### 硬件中断 CPU 列表 {#inputs.ebpf.network.nic_optimize.irq_cpu_list}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.irq_cpu_list`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - irq_cpu_list: ''
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | string |
+
+**详细描述**:
+
+用于处理网卡硬件中断的 CPU ID 或 CPU 列表。
+
+数量建议与 RSS 队列数量一致。
+若启用 XDP CPU Redirect，仅需指定一个 CPU。
+
+可设置为：
+  - 指定 CPU 列表（如 2,4,6）
+  - local（自动匹配本地 NUMA 节点 CPU）
+
+建议所选 CPU 与物理网卡位于同一 NUMA 节点。
+
+##### 启用 XDP CPU 重定向 {#inputs.ebpf.network.nic_optimize.xdp_cpu_redirect}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.xdp_cpu_redirect`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - xdp_cpu_redirect: false
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | bool |
+
+**详细描述**:
+
+是否启用 XDP CPUMAP 重定向分发。
+
+用于解决硬件 RSS 无法对封装报文
+（如 Double VLAN、ERSPAN 等）进行均匀分摊，
+导致单核过载和丢包的问题。
+
+##### XDP 队列大小 {#inputs.ebpf.network.nic_optimize.xdp_queue_size}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.xdp_queue_size`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - xdp_queue_size: 2048
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | int |
+
+**详细描述**:
+
+XDP CPUMAP 队列大小。
+
+取值范围：[512, 8192]，建议配置为 2 的幂。
+
+增大可提升突发流量承载能力，但会占用更多内存。
+
+##### XDP 重定向 CPU 列表 {#inputs.ebpf.network.nic_optimize.xdp_cpu_redirect_list}
+
+**标签**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.xdp_cpu_redirect_list`
+
+**默认值**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - xdp_cpu_redirect_list: ''
+```
+
+**模式**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | string |
+
+**详细描述**:
+
+XDP 重定向后用于处理数据包的 CPU 列表。
+
+填写样式如：4,6,8
+
 ### 调优 {#inputs.ebpf.tunning}
 
 #### 采集队列大小 {#inputs.ebpf.tunning.collector_queue_size}
diff --git a/server/agent_config/README.md b/server/agent_config/README.md
index f1329ba5f04..3d3dc7b5b44 100644
--- a/server/agent_config/README.md
+++ b/server/agent_config/README.md
@@ -5304,6 +5304,372 @@ inputs:
 Disable Node.js (V8) interpreter profiling. When disabled, Node.js process stack traces will not be collected,
 saving approximately 6.4 MB of kernel memory (v8_unwind_info_map).
 
+### Network {#inputs.ebpf.network}
+
+#### NIC optimization Enabled {#inputs.ebpf.network.nic_opt_enabled}
+
+**Tags**:
+
+`hot_update`
+<mark>ee_feature</mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_opt_enabled`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_opt_enabled: false
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | bool |
+
+**Description**:
+
+Whether to enable NIC optimization for enhanced multi-core packet
+processing and burst traffic resilience.
+
+When enabled, the system applies a combination of:
+  - RSS hardware queue configuration
+  - RX ring descriptor size tuning
+  - IRQ (interrupt) CPU affinity binding
+  - Optional XDP CPUMAP-based CPU redirection
+
+This optimization mitigates scenarios where RSS hardware cannot hash
+inner headers of encapsulated traffic (e.g., GRE, Double VLAN,
+VXLAN, ERSPAN), which may otherwise cause traffic to be concentrated
+on a single CPU core and lead to packet drops or performance bottlenecks.
+
+RX ring tuning improves burst handling capability by increasing
+the number of descriptors available for packet reception, reducing
+the likelihood of ring overflow under high traffic conditions.
+
+When XDP CPU redirect is enabled, packets are redistributed in
+software across multiple CPU cores after initial reception,
+providing better load balancing beyond hardware RSS capabilities.
+
+Recommended to enable this feature when:
+  1) Traffic on the interface consists primarily of encapsulated
+     packets (e.g., verified via tcpdump showing GRE, Double VLAN,
+     VXLAN, etc.).
+  2) One CPU core shows near 100% softirq utilization (e.g.,
+     observed via `top` with per-CPU view), while other CPUs
+     remain underutilized.
+
+For optimal performance, IRQ CPUs and XDP redirect CPUs should be
+configured on the same NUMA node as the physical NIC.
+
+#### NIC Optimize {#inputs.ebpf.network.nic_optimize}
+
+**Tags**:
+
+`hot_update`
+<mark>ee_feature</mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - interface: ''
+        irq_cpu_list: ''
+        rss_channel_count: 0
+        rx_ring_size: 0
+        xdp_cpu_redirect: false
+        xdp_cpu_redirect_list: ''
+        xdp_queue_size: 2048
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | dict |
+
+**Description**:
+
+Configure NIC-level performance optimizations for specific interfaces.
+
+This feature improves packet processing scalability and burst handling
+by tuning hardware RSS queues, interrupt CPU affinity, RX ring size,
+and optional XDP CPUMAP-based CPU redirection.
+
+Recommended when:
+  - Traffic is primarily encapsulated (GRE, Double VLAN, VXLAN, ERSPAN).
+  - One CPU shows near 100% softirq usage while others are idle.
+
+To achieve better performance, the program will automatically disable the
+irqbalance service to prevent network interface interrupts from migrating
+between CPUs.
+
+Multiple NIC optimize entries can be configured for different interfaces.
+
+Example:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_opt_enabled: true
+      nic_optimize:
+      - interface: eth0
+        rx_ring_size: 4096
+        rss_channel_count: 2
+        irq_cpu_list: 1,2
+        xdp_cpu_redirect: true
+        xdp_queue_size: 2048
+        xdp_cpu_redirect_list: 4,5,6,7
+      - interface: eth1
+        rx_ring_size: 4096
+        rss_channel_count: 2
+        irq_cpu_list: 1,2
+        xdp_cpu_redirect: true
+        xdp_queue_size: 2048
+        xdp_cpu_redirect_list: 4,5,6,7
+```
+
+##### Interface {#inputs.ebpf.network.nic_optimize.interface}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.interface`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - interface: ''
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | string |
+
+**Description**:
+
+Name of the network interface to optimize.
+
+##### RX Ring Size {#inputs.ebpf.network.nic_optimize.rx_ring_size}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.rx_ring_size`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - rx_ring_size: 0
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | int |
+
+**Description**:
+
+Number of RX descriptors in NIC receive ring.
+
+Increasing this value improves burst traffic buffering
+and reduces packet drops caused by ring overflow.
+Specifically, use `ethtool -g <iface>` to check the current
+configuration, and adjust to an appropriate value based on your workload.
+
+0 (default) means keep the original state and ignore this setting.
+
+##### RSS Channel Count {#inputs.ebpf.network.nic_optimize.rss_channel_count}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.rss_channel_count`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - rss_channel_count: 0
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | int |
+
+**Description**:
+
+Number of hardware RSS queues.
+Determines how many queues packets are distributed to after
+hardware hash calculation.
+
+Maximum supported value is typically 16 and must not exceed
+the number of logical CPU cores.
+Specifically, use `ethtool -l <iface>` to check the current configuration
+and adjust to an appropriate value based on your workload.
+
+When XDP CPU redirect is enabled, it is recommended to set this to 1.
+0 (default) means keep the original state and ignore this setting.
+
+##### Hardware IRQ CPU List {#inputs.ebpf.network.nic_optimize.irq_cpu_list}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.irq_cpu_list`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - irq_cpu_list: ''
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | string |
+
+**Description**:
+
+CPU ID or comma-separated CPU list used for handling NIC interrupts.
+
+Recommended to match the number of RSS queues.
+If XDP CPU redirect is enabled, only one CPU is required.
+
+Value can be:
+  - Specific CPU list (e.g., 2,4,6)
+  - "local" (auto match CPUs in local NUMA node)
+
+CPUs should be located on the same NUMA node as the NIC.
+
+##### Enable XDP CPU Redirect {#inputs.ebpf.network.nic_optimize.xdp_cpu_redirect}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.xdp_cpu_redirect`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - xdp_cpu_redirect: false
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | bool |
+
+**Description**:
+
+Enable XDP CPUMAP redirect to redistribute packets across CPUs
+in software.
+
+Useful when hardware RSS cannot distribute encapsulated traffic
+(e.g., Double VLAN, ERSPAN) evenly across CPUs, resulting in
+single-core overload and packet drops.
+
+##### XDP Queue Size {#inputs.ebpf.network.nic_optimize.xdp_queue_size}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.xdp_queue_size`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - xdp_queue_size: 2048
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | int |
+
+**Description**:
+
+Size of the XDP CPUMAP queue.
+
+Valid range: [512, 8192]. Powers of two are recommended.
+
+Larger values improve burst tolerance but consume more memory.
+
+##### XDP Redirect CPU List {#inputs.ebpf.network.nic_optimize.xdp_cpu_redirect_list}
+
+**Tags**:
+
+<mark></mark>
+
+**FQCN**:
+
+`inputs.ebpf.network.nic_optimize.xdp_cpu_redirect_list`
+
+**Default value**:
+```yaml
+inputs:
+  ebpf:
+    network:
+      nic_optimize:
+      - xdp_cpu_redirect_list: ''
+```
+
+**Schema**:
+| Key  | Value                        |
+| ---- | ---------------------------- |
+| Type | string |
+
+**Description**:
+
+CPU list used for processing packets after XDP redirection.
+
+Format example: 4,6,8
+
 ### Tunning {#inputs.ebpf.tunning}
 
 #### Collector Queue Size {#inputs.ebpf.tunning.collector_queue_size}
diff --git a/server/agent_config/template.yaml b/server/agent_config/template.yaml
index 1095afc4188..03547d208b9 100644
--- a/server/agent_config/template.yaml
+++ b/server/agent_config/template.yaml
@@ -3870,6 +3870,342 @@ inputs:
         nodejs_disabled: false
     # type: section
     # name:
+    #   en: Network
+    #   ch: 网络
+    # description:
+    network:
+      # type: bool
+      # name: NIC optimization Enabled
+      # unit:
+      # range: []
+      # enum_options: []
+      # modification: hot_update
+      # ee_feature: true
+      # description:
+      #   en: |-
+      #     Whether to enable NIC optimization for enhanced multi-core packet
+      #     processing and burst traffic resilience.
+      #
+      #     When enabled, the system applies a combination of:
+      #       - RSS hardware queue configuration
+      #       - RX ring descriptor size tuning
+      #       - IRQ (interrupt) CPU affinity binding
+      #       - Optional XDP CPUMAP-based CPU redirection
+      #
+      #     This optimization mitigates scenarios where RSS hardware cannot hash
+      #     inner headers of encapsulated traffic (e.g., GRE, Double VLAN,
+      #     VXLAN, ERSPAN), which may otherwise cause traffic to be concentrated
+      #     on a single CPU core and lead to packet drops or performance bottlenecks.
+      #
+      #     RX ring tuning improves burst handling capability by increasing
+      #     the number of descriptors available for packet reception, reducing
+      #     the likelihood of ring overflow under high traffic conditions.
+      #
+      #     When XDP CPU redirect is enabled, packets are redistributed in
+      #     software across multiple CPU cores after initial reception,
+      #     providing better load balancing beyond hardware RSS capabilities.
+      #
+      #     Recommended to enable this feature when:
+      #       1) Traffic on the interface consists primarily of encapsulated
+      #          packets (e.g., verified via tcpdump showing GRE, Double VLAN,
+      #          VXLAN, etc.).
+      #       2) One CPU core shows near 100% softirq utilization (e.g.,
+      #          observed via `top` with per-CPU view), while other CPUs
+      #          remain underutilized.
+      #
+      #     For optimal performance, IRQ CPUs and XDP redirect CPUs should be
+      #     configured on the same NUMA node as the physical NIC.
+      #
+      #   ch: |-
+      #     是否启用网卡优化功能，用于提升多核环境下的网络包处理能力
+      #     以及突发流量承载能力。
+      #
+      #     开启后将综合进行以下优化：
+      #       - RSS 硬件队列数量配置
+      #       - RX ring 描述符数量调优
+      #       - 硬件中断与 CPU 亲和性绑定（IRQ 绑核）
+      #       - 可选的 XDP CPUMAP 软件重定向分发
+      #
+      #     该优化主要解决 RSS 硬件无法基于封装报文内层头部
+      #     （如 GRE、Double VLAN、VXLAN、ERSPAN）进行哈希分摊的问题，
+      #     避免流量集中在单个 CPU 上造成过载和丢包。
+      #
+      #     通过调整 RX ring 描述符数量，可提升突发流量场景下的
+      #     接收缓存能力，降低 ring 满导致的丢包风险。
+      #
+      #     在启用 XDP CPU Redirect 时，数据包会在接收后通过
+      #     CPUMAP 在多个 CPU 核心之间进行软件层重分发，
+      #     从而突破硬件 RSS 的能力限制，实现更均衡的负载分布。
+      #
+      #     建议在以下场景开启该功能：
+      #       1）使用 tcpdump 抓包发现该接口流量主要为 GRE、
+      #          Double VLAN、VXLAN 等封装报文；
+      #       2）通过 `top`（按 1）观察各 CPU 软中断占用率时，
+      #          发现某一个 CPU 的 softirq 接近 100%，
+      #          而其他 CPU 软中断占用率明显偏低。
+      #
+      #     为获得最佳性能，建议将中断 CPU 和 XDP 重定向 CPU
+      #     配置在与物理网卡相同的 NUMA 节点上。
+      nic_opt_enabled: false
+      # type: dict
+      # name:
+      #   en: NIC Optimize
+      #   ch: 网卡优化配置
+      # unit:
+      # range: []
+      # enum_options: []
+      # modification: hot_update
+      # ee_feature: true
+      # description:
+      #   en: |-
+      #     Configure NIC-level performance optimizations for specific interfaces.
+      #
+      #     This feature improves packet processing scalability and burst handling
+      #     by tuning hardware RSS queues, interrupt CPU affinity, RX ring size,
+      #     and optional XDP CPUMAP-based CPU redirection.
+      #
+      #     Recommended when:
+      #       - Traffic is primarily encapsulated (GRE, Double VLAN, VXLAN, ERSPAN).
+      #       - One CPU shows near 100% softirq usage while others are idle.
+      #
+      #     To achieve better performance, the program will automatically disable the
+      #     irqbalance service to prevent network interface interrupts from migrating
+      #     between CPUs.
+      #
+      #     Multiple NIC optimize entries can be configured for different interfaces.
+      #
+      #     Example:
+      #     ```yaml
+      #     inputs:
+      #       ebpf:
+      #         network:
+      #           nic_opt_enabled: true
+      #           nic_optimize:
+      #           - interface: eth0
+      #             rx_ring_size: 4096
+      #             rss_channel_count: 2
+      #             irq_cpu_list: 1,2
+      #             xdp_cpu_redirect: true
+      #             xdp_queue_size: 2048
+      #             xdp_cpu_redirect_list: 4,5,6,7
+      #           - interface: eth1
+      #             rx_ring_size: 4096
+      #             rss_channel_count: 2
+      #             irq_cpu_list: 1,2
+      #             xdp_cpu_redirect: true
+      #             xdp_queue_size: 2048
+      #             xdp_cpu_redirect_list: 4,5,6,7
+      #     ```
+      #
+      #   ch: |-
+      #     针对指定网卡接口进行性能优化配置。
+      #
+      #     该功能通过调优 RSS 队列、中断绑核、RX ring 大小、
+      #     以及可选的 XDP CPUMAP 重定向机制，提升多核扩展能力
+      #     和突发流量承载能力。
+      #
+      #     建议在以下场景开启：
+      #       - 接口流量主要为 GRE、Double VLAN、VXLAN、ERSPAN 等封装报文；
+      #       - 某个 CPU 的 softirq 占用率接近 100%，而其他 CPU 空闲。
+      #
+      #     为了获得更好的性能，程序会自动停用 irqbalance 服务，以防止网卡中断在 CPU 之间迁移。
+      #
+      #     可为多个接口分别配置优化策略。
+      #
+      #     样例:
+      #     ```yaml
+      #     inputs:
+      #       ebpf:
+      #         network:
+      #           nic_opt_enabled: true
+      #           nic_optimize:
+      #           - interface: eth0
+      #             rx_ring_size: 4096
+      #             rss_channel_count: 2
+      #             irq_cpu_list: 1,2
+      #             xdp_cpu_redirect: true
+      #             xdp_queue_size: 2048
+      #             xdp_cpu_redirect_list: 4,5,6,7
+      #           - interface: eth1
+      #             rx_ring_size: 4096
+      #             rss_channel_count: 2
+      #             irq_cpu_list: 1,2
+      #             xdp_cpu_redirect: true
+      #             xdp_queue_size: 2048
+      #             xdp_cpu_redirect_list: 4,5,6,7
+      #     ```
+      # upgrade_from:
+      # ---
+      # type: string
+      # name:
+      #   en: Interface
+      #   ch: 网卡接口
+      # description:
+      #   en: |-
+      #     Name of the network interface to optimize.
+      #
+      #   ch: |-
+      #     需要进行优化的网卡接口名称。
+      # ---
+      # interface: ""
+      # ---
+      # type: int
+      # name:
+      #   en: RX Ring Size
+      #   ch: RX Ring 描述符数量
+      # description:
+      #   en: |-
+      #     Number of RX descriptors in NIC receive ring.
+      #
+      #     Increasing this value improves burst traffic buffering
+      #     and reduces packet drops caused by ring overflow.
+      #     Specifically, use `ethtool -g <iface>` to check the current
+      #     configuration, and adjust to an appropriate value based on your workload.
+      #
+      #     0 (default) means keep the original state and ignore this setting.
+      #
+      #   ch: |-
+      #     网卡接收环（RX ring）的描述符数量。
+      #
+      #     增大该值可提升突发流量场景下的缓存能力，
+      #     降低因 ring 满导致的丢包风险。
+      #     具体使用`ethtool -g <iface>` 查看当前配置，根据实际情况调整到合适的值。
+      #
+      #     默认值为 0 表示保持原状忽略此项配置。
+      # ---
+      # rx_ring_size: 0
+      # ---
+      # type: int
+      # name:
+      #   en: RSS Channel Count
+      #   ch: RSS 队列数量
+      # description:
+      #   en: |-
+      #     Number of hardware RSS queues.
+      #     Determines how many queues packets are distributed to after
+      #     hardware hash calculation.
+      #
+      #     Maximum supported value is typically 16 and must not exceed
+      #     the number of logical CPU cores.
+      #     Specifically, use `ethtool -l <iface>` to check the current configuration
+      #     and adjust to an appropriate value based on your workload.
+      #
+      #     When XDP CPU redirect is enabled, it is recommended to set this to 1.
+      #     0 (default) means keep the original state and ignore this setting.
+      #
+      #   ch: |-
+      #     RSS 硬件队列数量。
+      #     数据包在物理网卡完成哈希后，将分发到指定数量的队列并触发中断。
+      #
+      #     最大一般支持 16，且不要超过逻辑 CPU 核数。
+      #     具体使用`ethtool -l <iface>` 查看当前配置，根据实际情况调整到合适的值。
+      #
+      #     当启用 XDP CPU Redirect 时建议设置为 1。
+      #     默认值为 0 表示保持原状忽略此项配置。
+      # ---
+      # rss_channel_count: 0
+      # ---
+      # type: string
+      # name:
+      #   en: Hardware IRQ CPU List
+      #   ch: 硬件中断 CPU 列表
+      # description:
+      #   en: |-
+      #     CPU ID or comma-separated CPU list used for handling NIC interrupts.
+      #
+      #     Recommended to match the number of RSS queues.
+      #     If XDP CPU redirect is enabled, only one CPU is required.
+      #
+      #     Value can be:
+      #       - Specific CPU list (e.g., 2,4,6)
+      #       - "local" (auto match CPUs in local NUMA node)
+      #
+      #     CPUs should be located on the same NUMA node as the NIC.
+      #
+      #   ch: |-
+      #     用于处理网卡硬件中断的 CPU ID 或 CPU 列表。
+      #
+      #     数量建议与 RSS 队列数量一致。
+      #     若启用 XDP CPU Redirect，仅需指定一个 CPU。
+      #
+      #     可设置为：
+      #       - 指定 CPU 列表（如 2,4,6）
+      #       - local（自动匹配本地 NUMA 节点 CPU）
+      #
+      #     建议所选 CPU 与物理网卡位于同一 NUMA 节点。
+      # ---
+      # irq_cpu_list: ""
+      # ---
+      # type: bool
+      # name:
+      #   en: Enable XDP CPU Redirect
+      #   ch: 启用 XDP CPU 重定向
+      # description:
+      #   en: |-
+      #     Enable XDP CPUMAP redirect to redistribute packets across CPUs
+      #     in software.
+      #
+      #     Useful when hardware RSS cannot distribute encapsulated traffic
+      #     (e.g., Double VLAN, ERSPAN) evenly across CPUs, resulting in
+      #     single-core overload and packet drops.
+      #
+      #   ch: |-
+      #     是否启用 XDP CPUMAP 重定向分发。
+      #
+      #     用于解决硬件 RSS 无法对封装报文
+      #     （如 Double VLAN、ERSPAN 等）进行均匀分摊，
+      #     导致单核过载和丢包的问题。
+      # ---
+      # xdp_cpu_redirect: false
+      # ---
+      # type: int
+      # name:
+      #   en: XDP Queue Size
+      #   ch: XDP 队列大小
+      # description:
+      #   en: |-
+      #     Size of the XDP CPUMAP queue.
+      #
+      #     Valid range: [512, 8192]. Powers of two are recommended.
+      #
+      #     Larger values improve burst tolerance but consume more memory.
+      #
+      #   ch: |-
+      #     XDP CPUMAP 队列大小。
+      #
+      #     取值范围：[512, 8192]，建议配置为 2 的幂。
+      #
+      #     增大可提升突发流量承载能力，但会占用更多内存。
+      # ---
+      # xdp_queue_size: 2048
+      # ---
+      # type: string
+      # name:
+      #   en: XDP Redirect CPU List
+      #   ch: XDP 重定向 CPU 列表
+      # description:
+      #   en: |-
+      #     CPU list used for processing packets after XDP redirection.
+      #
+      #     Format example: 4,6,8
+      #
+      #   ch: |-
+      #     XDP 重定向后用于处理数据包的 CPU 列表。
+      #
+      #     填写样式如：4,6,8
+      # ---
+      # xdp_cpu_redirect_list: ""
+      #
+      nic_optimize:
+        - interface: ""
+          rx_ring_size: 0
+          rss_channel_count: 0
+          irq_cpu_list: ""
+          xdp_cpu_redirect: false
+          xdp_queue_size: 2048
+          xdp_cpu_redirect_list: ""
+    # type: section
+    # name:
     #   en: Tunning
     #   ch: 调优
     # description: