HP delivers new 3000 performance advice

N-Class, Express 1 change picture in HP World talk

By Craig L. Solomon

As the financial community waited to hear how Alan Greenspan would change interest rates, the e3000 community waited at HP World to hear numbers: the latest results from HP’s e3000 performance expert Kevin Cooper.

The numbers have changed my thinking. Instead of dealing with systems I am thoroughly familiar with, I’ve had to enter a new world of HP e3000 systems, some of which have yet to hit the street. Benchmarks from the new N-Class four-way box were a dominating topic of conversation at Cooper’s talk and the HP World show in general. With the exception of a few beta locations and HP itself, none of us attending Cooper’s talk had the opportunity yet to see what could be the largest jump in history for e3000 computing power.

Cooper said HP is now basing their performance numbers on a series of tests showing straight processing power as well as batch throughput and completion times. According to Cooper, HP tests real-world performance by using “real customer applications and data. The HP e3000 is loaded to 95-percent busy on average, using a separate driver system to simulate online activity. Throughput is measured by the number of transactions completed during the 15 minutes of stable execution at this level.” Cooper also let us know that batch processes, although run separately, are also put through the same rigorous test.

When put through a series of CPU-intensive batch job tests, the top of the line N4000-400-550 completed them between one-half and one-third of their run time on a fully loaded 997. This represents a gain of 100-200 percent in batch performance.

Cooper showed how this effects real-time results by presenting the table below. These results show the time needed to sort an 800Mb file (10 million records of 80 bytes each).

System	Completion Time
997	13 minutes
989/x00	10 minutes
989/x650	8 minutes
N4000-440	5 minutes
N4000-550	4 minutes

“Using real customer data, a 40Gb Automatic Master separated into eight jumbo chunks,” Cooper said. “According to the customer, it took this site approximately 28 hours to complete a 25-percent capacity change. Using the N4000-550, this capacity change took 7 hours and 19 minutes.”

The performance numbers can be used as motivation for upgrades. With the demise of the beloved 9x7 line, some sites will choose to upgrade to stay on HP support, as well as take advantage of the benefits of MPE/iX 7.0 and later releases. I could see an upgrade path in the numbers Cooper presented in his talk, and I’ve reorganized his figures into the chart shown in Figure 1.

Cooper also used his talk to modify the rules for memory by introducing new recommendations for both A-Class and N-Class models. He said these are starting points for memory, and actual memory requirements may be higher.

• A-Class: 256Mb per CPU, plus application requirements.

• N-Class (220 and 330 Models): 512Mb per CPU, plus application requirements.

• N-Class (440 and 550 Models): 1Gb per CPU, plus application requirements.

Although HP suggests between one-half megabyte to 1Mb additional per user, I feel safer recommending 2Mb per user. If you use memory-intensive applications such as 4GL tools, have a very heavy batch environment, higher than average online users, or if you are going to be adding more processors, you may need to add more memory than listed above.

Express 1 improvements

Cooper’s performance talk, although predominately spent covering new hardware, did touch on changes surrounding the release of 7.0 Express 1 and subsequent patches.

On the high-end systems running 7.0 Express 1, HP has finally addressed the problem of limited PIN numbers. Hewlett-Packard has introduced the “BIGPIN” (8K to 12K) option within SYSGEN. By default, this option is off, but a site running a new N-Class system on 7.0 Express 1 can turn this option on, breaking the 8,190 pin barrier. Cooper advised using this with caution, since certain third-party applications as well as homegrown applications may not recognize a five-digit PIN.

HP e3000 Java fans will be glad to hear that Express 1 offers improved threads performance, resulting in improved Java performance. There have also been performance increases in the Java Virtual Machine.

As of June, two patches have been released which may improve performance on some larger systems running MPE/iX 6.5 and 7.0. The MPELXH8 patch for the e3000’s Memory Manager:

• Provides a new “Make_absent” option so systems programs and privileged mode third-party tools can explicitly free memory pages they no longer need.

• Reduces how often memory manager tries to proactively make free pages.

• Reduces the number of pages the memory manager tried to make free in each call.

• Reduces background overhead when closed files are mapped out of memory.

• Throttles back memory manager IOs by not flushing out dirty pages when the IO system is busy. (Note: Later patch MPELXM5 makes the same change at an additional code location in memory manager.)

• Reduces unnecessary overhead during Transaction Manager checkpoints, caused by a side effect from an earlier 6.5 patch (MPELX75 – fixed a system hang).

• Includes the “Best Of” from earlier 6.5 patches developed last year, MPELX66 in 6.5 PowerPatch 1 and MPELXB0 in PP2.

• Collects all 6.5 memory manager patches into one “good” patch for both 6.5 and 7.0.

A second patch, MPELXH3 for TurboSTORE has these effects:

• After storing a file, instructs the memory manager to immediately free up all pages of that file remaining in memory.

• Uses the new “make_absent” option provided in MPELXH8.

Cooper said these patches are not included in MPE/iX 6.5 Express 2 or the new MPE/iX 7.0 Express 1.

Craig L. Solomon is the founder of the IT Consulting Consortium, and can be reached at craig@craigs.com or at his Web site, www.craigs.com.