傳統的TCP/IP協定架構中，TCP提供可靠性的服務給上層，因此接收端的TCP保證送交上層的資料具有完整性及正確的順序。為了達到此目的，TCP收到的封包必須先經過暫存甚至重組的步驟，才能放進應用程式的記憶體空間裡。在高速的網路環境下，TCP內部的重組及複製資料的動作會浪費記憶體匯流排的頻寬、增加CPU處理資料的時間、需要額外的記憶體空間來放置資料，以及浪費多餘的時間在資料的搬移上，因而使得網路處理的效能大受影響。新訂定的iWARP協定組的主要功能便是支援遠端DMA存取（Remote DMA, RDMA）以及直接資料搬移（Direct Data Placement, DDP），可以把收到的資料直接放到最終將被使用的記憶體空間中，並利用此空間進行資料重組，以取代原本TCP內部所需的重組緩衝區，進而消除不必要的資料搬移。

本篇論文中提出一個實現iWARP協定組的架構，並探討此架構與現有TCP/IP卸載引擎（TCP/IP Offload Engine, TOE）的整合方式。此設計以實現標頭處理的硬體為主，可支援8條連線同時運作。使用UMC 0.18 的技術，目前處理速度可達到1-Gbps的乙太網路需求，並使用了139K個邏輯閘。此外，我們模擬出一個TOE的環境，實驗結果說明若是接收到的封包有亂序的情形發生時，整合了iWARP協定組的TOE可以大大減少網路卡上記憶體的使用量以及每個封包的平均等待時間。

In the architecture of traditional TCP/IP stack, TCP provides the upper layer with reliability. TCP at the receiving side must ensure the integrity and in-order of the data before it can be delivered to the host. Thus received TCP packets must be temporarily stored and even reassembled before they are submitted to the application’s memory space. In the environment of high-speed network, reassembly and data copy involve high demands on memory bus bandwidth, increases on the CPU processing overhead, extra memory space to store packets, and redundant time to move the data. The performance of network processing is therefore getting worse. The main features of the newly-proposed iWARP protocol suite are supporting remote DMA (RDMA) and direct data placement (DDP). Thus the received data can be directly placed into the destination memory space of application where they can be reassembled as well, so that the reassembly buffer required in TCP layer and unnecessary data copies can be removed.

We propose an architecture to implement the iWARP protocol suite in this thesis, and describe how we can integrate the design with the existing TCP/IP offload engine (TOE). Our design is a state machine based hardware that focuses on the header processing, and currently supports up to 8 connections at the same time. Using the technology of UMC 0.18, our design can meet the wire speed of 1-Gbps Ethernet, requiring about 139K gate count. In addition, we simulate a TOE environment, and the results of experiments show that the TOE integrated with our iWARP design can drastically reduce the memory requirement and average queuing time per packet on the network interface card (NIC).

[1] A. P. Foong, T. R. Huff, H. H. Hum, J. P. Patwardhan, and G. J. Regnier, “TCP
performance re-visited,” in Proc. ISPASS’03, pp. 70-79, 2003.
[2] J. M. Sol-Sloan, and I. Couvertier, “A parallel TCP/IP offload engine framework
for a TOE device,” in Proc. Applied Telecommunication Symposium, 2003.
[3] Y. Hoskote et al., “A TCP offload accelerator for 10 Gb/s Ethernet in 90-nm
CMOS,” in Proc. JSSC’03, vol. 38, NO. 11, pp. 1866-1875, Nov. 2003.
[4] K. Kant, “TCP offload performance for front-end servers,” in Proc.
GLOBECOM’03, pp. 3242-3247, 2003.
[5] Li Zhao, R. Illikkal, S. Makineni, and L. Bhuyan, “TCP/IP cache characterization in
commercial server workloads,” in Workshop on CAECW-7, Feb. 2004.
[6] S. Makineni and R. Iyer, “Architectural characterization of TCP/IP packet processing
on the Pentium M microprocessor,” in Proc. HPCA’04, p.152, Feb. 2004.
[7] R. Recio et al., “A Remote Direct Memory Access Protocol Specification,” Internet
Draft, Remote Direct Data Placement Work Group, Sept. 8, 2006. Available:
http://www.ietf.org/internet-drafts/draft-ietf-rddp-rdmap-07.txt
[8] H. Shah et al., “Direct Data Placement over Reliable Transports,” Internet Draft,
Remote Direct Data Placement Work Group, Sept. 9, 2006. Available:
http://www.ietf.org/internet-drafts/draft-ietf-rddp-ddp-07.txt
[9] P. Culley et al., “Marker PDU Aligned Framing for TCP Specification,” Internet
Draft, Remote Direct Data Placement Work Group, Oct. 7, 2006. Available:
http://www.ietf.org/internet-drafts/draft-ietf-rddp-mpa-08.txt
[10] B. Hausauer, “iWARP Ethernet: eliminating overhead in data center designs.”
Available: http://www.analogzone.com/nett0403.pdf
[11] C. B. Reardon, A. D. George, and C. T. Cole, “Comparative performance analysis
of RDMA-enhanced Ethernet,” in Proc. HPIDC’05, July 2005.
[12] P. Balaji, H. V. Shah, and D. K. Panda, “Sockets vs RDMA interface over
10-Gigabit networks: an in-depth analysis of the memory traffic bottleneck," in
Workshop on RAIT, Sept. 2004.
[13] P-K. Chan, C-H Chen, and C-Y Yu, “An iWARP-based TCP/IP offload engine,” in
Proc. 17th VLSI Design/CAD Symposium, 2006.
[14] M. J. Rashti and A. Afsahi, “NetEffect PCI-Express 10-Gigabit iWARP Ethernet: a
performance study,” product’s white paper. Available:
http://www.neteffect.com/documents/Queens_University_whitepaper.pdf
[15] P. Balaji, H. –W. Jin, K. Vaidyanathan, and D. K. Panda, “Supporting iWARP
80
compatibility and features for regular network adapters,” in Proc. RAIT’05, Sept.
2005.
[16] D. Dalessandro, A. Devulapalli, and P. Wyckoff, “Design and implementation of the
iWARP protocol in software,” in Proc. PDCS’05, pp. 471-476, Nov. 2005.
[17] Ammasso Inc. “Ammasso 1100.” URL: http://www.ammasso.com/products.htm
[18] D. Dalessandro, P. Wyckoff, and A. Devulapalli, OSC iWARP team. URL:
http://www.osc.edu/research/network_file/projects/iwarp/iwarp_main.shtml
[19] Lamprey Networks, Inc. URL:
http://lampreynetworks.com/Content/Test_Suites/testsuite_iWARP_overview.html
[20] Mentor Graphics Corporation. URL: http://www.model.com/
[21] A. Morton et al., “Packet Reordering Metrics,” RFC 4737, Nov. 2006.
[22] A. Earls, “TCP offload engines finally arrive,” Storage Magazine, Mar. 2002, URL:
http://www.iscsistorage.com/iscsidevices.htm
[23] InfiniBand Trade Association. URL: http://www.infinibandta.org/home
[24] URL: http://www.ethereal.com/