on60 transactions per second before running out on the atestnd reat the c nee d to string together a cluster of at le st 10 servers to reach and maint on in te has in our competitors to pitch tation the eaper solution.The same hardware.It is often the case that the most efficient solution wins the bid. The limits imposed by physics might soo ut the brakes on the fantastic growth of pr sor speed [Lewl].Not so for eabits per second is common.The end of the road for communication speed is nowhere in sight [Lew2] per-second all-opti cal networking.The biggest obstacle currently is not chnical nature:it is the y fron faste than the computing dev to them.Emerging network te gies such as 100 Mbps LAN e has Po ith 1 E g adapt leaving software perforr ce bottlenecks undetected.Not so ,000 instructions I to the pro receive pat graded throughput on a Fast Eth net adapter. very fw computers are capable of saturating a becoming the new bottleneck,and it's going to stay that way. ake a lo bred of bandwith-and yc-hun y applications that push the boundaries of techn y burden on the sof vare.Further advances in execution speed will depend Terminology minimize the use of mem mory in a softy nted in terms of response time and ile time able size. oved the topic of space efficiency for its own sake to the back xii xii Imagine that you are trying to sell your product, say a Web application server, to a Fortune 500 company. They need 600 transactions per second to run their business online. Your application server can support only 60 transactions per second before running out of steam on the latest and greatest server hardware. If the customer is to use your software, they need to string together a cluster of at least 10 servers to reach their 600-transaction per second goal, raising the cost of your solution in terms of hardware, software licenses, network administration, and maintenance. To make matters worse, the customer has invited two of your competitors to pitch their own solutions. If a competitor has a more efficient implementation, they will need less hardware to deliver the required performance, and they will offer a cheaper solution. The speed of the processor is a constant in this situation—the software vendors in this story compete over the same hardware. It is often the case that the most efficient solution wins the bid. You also must examine how processing speed compares to communication speed. If we can transmit data faster than the computer can generate it, then the computer (processor plus software) is the new bottleneck. The limits imposed by physics might soon put the brakes on the fantastic growth of processor speed [Lew1]. Not so for communication speed. Like processing speed, communication speed has enjoyed phenomenal growth. Back in 1970, 4800 bits per second was considered high-speed communication. Today, hundreds of megabits per second is common. The end of the road for communication speed is nowhere in sight [Lew2]. Optical communication technology does not seem to have show-stopping technological roadblocks that will threaten progress in the near future. Several research labs are already experimenting with 100-gigabit￾per-second all-optical networking. The biggest obstacle currently is not of a technical nature; it is the infrastructure. High-speed networking necessitates the rewiring of the information society from copper cables to optical fiber. This campaign is already underway. Communication adapters are already faster than the computing devices attached to them. Emerging network technologies such as 100 Mbps LAN adapters and high-speed ATM switches make computer speed critical. In the past, inefficient software has been masked by slow links. Popular communication protocols such as SNA and TCP/IP could easily overwhelm a 16 Mbps token ring adapter, leaving software performance bottlenecks undetected. Not so with 100 Mbps FDDI or Fast Ethernet. If 1,000 instructions are added to the protocol's send/receive path, they may not degrade throughput on a token ring connection because the protocol implementation can still pump data faster than the token ring can consume it. But an extra 1,000 instructions show up instantly as degraded throughput on a Fast Ethernet adapter. Today, very few computers are capable of saturating a high-speed link, and it is only going to get more difficult. Optical communication technology is now surpassing the growth rate of microprocessor speed. The computer (processor plus software) is quickly becoming the new bottleneck, and it's going to stay that way. To make a long story short, software performance is important and always will be. This one is not going away. As processor and communication technology march on, they redefine what "fast" means. They give rise to a new breed of bandwidth- and cycle-hungry applications that push the boundaries of technology. You never have enough horsepower. Software efficiency now becomes even more crucial than before. Whether the growth of processor speed is coming to an end or not, it will definitely trail communication speed. This puts the efficiency burden on the software. Further advances in execution speed will depend heavily on the efficiency of the software, not just the processor. Terminology Before moving on, here are a few words to clarify the terminology. "Performance" can stand for various metrics, the most common ones being space efficiency and time efficiency. Space efficiency seeks to minimize the use of memory in a software solution. Likewise, time efficiency seeks to minimize the use of processor cycles. Time efficiency is often represented in terms of response time and throughput. Other metrics include compile time and executable size. The rapidly falling price of memory has moved the topic of space efficiency for its own sake to the back burner. Desktop PCs with plenty of RAM (Random Access Memory) are common. Corporate customers