Xincuan 서버 | 엔터프라이즈 서버 하드웨어 공급업체
그만큼 40 Percent CPU Waste That No One Sees
Your server is processing 25 Gbps of network traffic. The CPU is at 60 활용률. The application is using 20 퍼센트. The other 40 percent is the kernel’s network stack — copying packets, computing checksums, segmenting TCP flows, and routing interrupts — all of which is work that a modern NIC can do in silicon. The gap between what your CPU does and what it could delegate to the adapter is the difference between a server that handles 25 Gbps and one that handles 100 Gbps. This article explains the offload technologies that move network processing from the CPU to the NIC — and when each one matters.
The Offload Technologies, Explained
| 기술 | What It Offloads | Supported On | CPU Savings | When It Matters |
|---|---|---|---|---|
| Checksum Offload (IP/TCP/UDP) | Computes IP/TCP/UDP checksums in the NIC instead of the CPU | Every modern NIC (Intel X710+, 브로드컴 57412+, 첼시 T6) | 2-5 percent CPU per 10 Gbps | Always enabled. Table stakes since 2005. |
| TSO/GSO (TCP Segmentation Offload) | NIC splits large TCP segments (64KB) into MTU-sized packets, generating all the headers in hardware | Every modern NIC | 10-20 percent CPU for large transfers | Big data transfers, 백업, media streams, VM migration |
| LRO/GRO (Large Receive Offload) | NIC merges incoming segments into large packets before the kernel processes them, reducing per-packet overhead | Every modern NIC | 10-15 percent CPU for receive-heavy workloads | Receive-heavy: file servers, 웹 서버, storage targets |
| RSS (Receive Side Scaling) | Distributes incoming packets across multiple CPU cores using a hash of the 5-tuple — parallelizes network processing | Every modern NIC (multiqueue) | Scales to N cores instead of 1 | Multi-core servers. Critical for 25GbE+ where a single core cannot process line rate. |
| RPS/RFS (Receive Packet Steering / Flow Steering) | Software complement to RSS. RPS balances load across cores; RFS steers packets to the core running the consuming application. | 리눅스 커널 (works with any NIC) | Improves cache locality, reduces latency | High-connection-count workloads (proxy, load balancer, 레디스) |
| 발가락 (TCP 오프로드 엔진) | Full TCP/IP protocol stack processed on the NIC. CPU never sees TCP state machine. | Only Chelsio T6 (and older proprietary stacks) | 까지 60 percent CPU savings on pure TCP workloads | Legacy/niche. Most NIC vendors dropped TOE due to kernel-netstack improvements. Chelsio keeps it for iWARP. |
| VXLAN/GENEVE Offload | NIC adds/removes overlay headers, computes overlay checksums, handles UDP encapsulation in hardware | Intel X710+, E810, Broadcom 574xx, 첼시 T6, NVIDIA ConnectX | 10-25 percent CPU on overlay networks | VXLAN environments (구름, SDN, NSX, OpenStack, Kubernetes CNI overlays) |
| SR-IOV (Single Root I/O Virtualization) | NIC presents virtual functions (VFs) directly to VMs/containers, bypassing the hypervisor’s network stack | 인텔, 브로드컴, 멜라녹스, Chelsio — all modern adapters | 까지 80 퍼센트 (VM networking bypass) | High-throughput virtualized workloads, DPDK applications, VNFs (NFV) |
| ADQ (애플리케이션 장치 대기열) | Dedicated NIC queues per application, zero-copy delivery from NIC to app memory | 인텔 E810 (exclusive) | 40-60 percent P99 latency reduction | Latency-sensitive apps: 레디스, Kafka, databases on Intel E810 |
What Actually Matters for Your Workloads
Virtualization and Hypervisors
VM웨어 ESXi, Hyper-V, and KVM rely on two offloads above all: RSS for multi-core parallelism and SR-IOV for VM networking. An ESXi host with 2x 25GbE NICs and RSS enabled distributes traffic across 8+ CPU cores, preventing single-core saturation. SR-IOV passes virtual functions directly to VMs, eliminating hypervisor packet processing — the difference between 20 Gbps and 35 Gbps of VM throughput on a 25 Gbps link.
Storage Networking (iSCSI, NVMe-oF, vSAN)
Storage traffic is TCP-based (iSCSI) or RDMA-based (NVMe-oF over RoCE/iWARP). Checksum and TSO offloads are essential for iSCSI — a 100 GB backup transfer without TSO offload consumes 10-20 percent of CPU that should be serving applications. NVMe-oF over RDMA bypasses the TCP stack entirely via RDMA offload, which is why RDMA-capable NICs (인텔 E810, 첼시 T6, Mellanox ConnectX) matter for NVMe-oF.
High-Connection-Count Applications (Proxy, Load Balancer, 레디스, Kafka)
These workloads generate many small packets and many connections. RSS + RPS/RFS keep the interrupt load balanced across cores. ADQ (인텔 E810) goes further with per-application queues that eliminate cross-application contention. 그만큼 40-60 percent P99 latency improvement from ADQ is the largest single latency win available from a NIC feature in 2026.
Virtualized Network Functions (NFV, DPDK)
VNFs (vRouters, vFirewalls, vLoad balancers) process millions of packets per second. SR-IOV passes VFs directly to the VM, and DPDK poll-mode drivers bypass the kernel entirely. ㅏ 25 Gbps NFV workload without SR-IOV runs at ~8 Gbps; with SR-IOV + DPDK it approaches 25 Gbps line rate. This is the 3x throughput difference that NFV vendors cite.
When Offloads Hurt — The Counterintuitive Cases
| 대본 | Offload That Hurts | Why | Fix |
|---|---|---|---|
| VMware ESXi with RSS on single-queue VMs | RSS (if misconfigured) | RSS hashes per-flow. A single large flow (one VM backup) lands on one core regardless of RSS. Multi-queue needed per VM. | Enable multi-queue virtio/VMXNET3, verify RSS on both host and VM sides |
| LRO on latency-sensitive OLTP | LRO/GRO | LRO merges packets, adding micro-buffering latency. OLTP needs per-packet delivery. | Disable LRO on database-facing NIC ports (keep on backup ports) |
| TSO with fragmented packets (NFS over TCP) | TSO | Fragmented NFS writes cannot be segmented by TSO; the NIC falls back to CPU segmentation, and the mixed path is slower than no TSO. | Keep TSO enabled; the fallback is rare. Monitor for NIC CPU usage anomalies. |
| TOE with modern kernel (kernel bypass) | 발가락 (if available) | Modern kernels have optimized netstacks; TOE conflicts with kernel socket APIs and firewalls. | Use RSS + 체크섬 + TSO instead. TOE is legacy. |
How to Verify Offloads Are Working
- 리눅스:
ethtool -k eth0shows current offload settings.ethtool -S eth0 | grep rssshows RSS queue distribution.nstat -azshows TCP retransmissions and checksum errors (indicators of offload issues). - 윈도우: Get-NetAdapterAdvancedProperty shows offload settings. Performance Monitor: Network Interface counters per adapter.
- VM웨어 ESXi: esxcli network nic get -n vmnic0 shows offload capabilities. esxtop network views show interrupt distribution across CPUs.
Source Your Networking Hardware Through Xincuan
We supply the full range of NICs — Intel X710, E810, 브로드컴 57412, 첼시 T6, Mellanox ConnectX — pre-installed in Dell PowerEdge and xFusion FusionServer servers with factory-direct pricing, 3-1년 보증, 그리고 글로벌 배송. Our engineers can help you configure offloads correctly for your workload.
Request a networking consultation and quote
탐구하다: Intel E810 25GbE | Chelsio T6 iWARP | Intel X710 10GbE