Reprinted by permission

FLEX-ESL - The New Kid On The Block

On September 2, 1998 Fundamental Software, Inc. (FSI) announced the availability of a S/390 class 'mainframe' computing solution that runs on a standard Intel server under Unix. The FSI FLEX-ESL system is based on the FLEX-ES ESA/370 'S/390 on Intel' emulator and when running under an Intel Pentium-II 400 MHz processor achieves approximately 7 to 8 MIPS of S/390 processing power. The FLEX-ES emulator allows you to use an IBM ESL license to run all the latest releases of IBM's mainframe operating systems including OS/390 2.6, VSE/ESA 2.3 and VM/ESA 2.3. Older releases of these operating systems are also fully supported.

The FLEX-ESL offering is based on a DELL server running dual Intel Pentium-II 400 processors, 128MB of memory and 2GB of RAID 1 DASD. The base system is very expandable and priced very competively.

S/390 On An Intel Platform

The FSI FLEX-ESL is based on a DELL PowerEdge 2300 server running two Intel Pentium-II 400 MHz processors. The first processor is dedicated to the emulator function and provides about 7 to 8 MIPS of processing power. The second processor is the I/O subsystem and runs under SCO Unixware 7. The first 128MB of memory on the server is reserved for SCO Unixware 7 and the FLEX-ES emulator. This leaves up to 896MB of memory for the mainframe operating system to use. The DELL PowerEdge 2300 server supports a dual channel RAID SCSI adapter providing large amounts of available DASD. FSI also has a three-channel Parallel Channel Adapter (PCA) for attaching 'bus and tag' devices.

Is ESA/370 Emulation Really a Restriction?

The FLEX-ES emulator creates a complete ESA/370 machine environment including some of the ESA/390 machine environment. What is missing from the ESA/390 machine?

Expanded Storage
Immediate and Relative Instructions
Hardware Data Compression
Cryptographic Instructions
Extended Configuration Mode
String Operations

As you can see the missing parts of the ESA/390 machine environment are not major and do not effect OS/390, VM/ESA or VSE/ESA to any great degree. In fact, these operating systems only require an ESA/370 machine environment.

Device Support

Device support on the FLEX-ESL is in many ways similar to the IBM PC Server 330 System/390 device support. The FLEX-ESL supports a wide range of both physical and emulated peripheral devices. These devices include emulated devices (DASD, etc.), 'Bus and Tag' channel attached devices, SCSI attached devices and integrated communication devices.

Emulated Devices

The FLEX-ESL emulates virtual all known types of FBA and CKD/ECKD DASD as well as many types of tape drives, terminals, printers, card readers and punches. The emulated devices are implemented in software running on the DELL server.

Parallel Channel Adapter (PCA)

The FLEX-ESL Parallel Channel Adapter provides three 'Bus and Tag' channels on a PCI adapter. Almost any type of 'Bus and Tag' device can be attached to the PCA with the exception of DASD.

Tape Devices

The FLEX-ESL provides support for many types of physical tape drives. SCSI attached 3420, 3480, 3490E, DAT and DLT drives are all supported and fully compatible with the IBM PC Server 330 System/390 tape drives. 'Bus and Tag' 3420, 3480, 3490E drives are also fully supported when attached to a PCA adapter card.

Communications

The FLEX-ESL provides support for various terminal and console devices. Devices can be attached through the use of the FLEX-ES Integrated Communications Adapter (ICA), PCA attachment of physical communication controllers or though the use of various emulated terminal and console facilities.

Shared Internal DASD

One of the most intriguing features of the FLEX-ES emulator is its ability to emulate a shared 3990 caching controller. Shared DASD? Yes. The FLEX-ES emulator allows two or more physically separate systems to share internal DASD over an Ethernet connection. Using a dedicated 100baseTX full duplex hub this connection provides a bi-directional transfer rate of up to 10MB per second between systems.

Shared Internal DASD Block Diagram

The 10MB per second transfer rate is equivalent to about 50% of the speed of an IBM ESCON channel connection. While sharing all internal DASD volumes of each system would not have good performance, sharing a limited number of internal DASD volumes will provide shared multiple system DASD access with excellent performance. Users running VSE/ESA can define a VSE Lock File and couple up to nine (9) separate FLEX-ESL systems together in a POWER Shared Spooling environment. A system configured in this manor could provide up to 70 MIPS of processing power spread across nine linked VSE/ESA systems.

Multiple Image Support

The FLEX-ES emulator also provides the ability to run multiple host images on a single FLEX-ESL system. The multiple image support of the FLEX-ES emulator is similar to partitioning a system using IBM's PR/SM. Using the FLEX-ES multiple image facility, it is easy to set up a second host image for software testing or installation of new software. Only available memory and CPU system resources limit the number of images.

FLEX-ESL Processor Speed

FSI claims that the FLEX-ESL system based on dual Pentium-II 400 processors provides the same CPU processing power as an IBM PC Server 330 System/390 with a P/390E adapter. It is well known that the P/390E adapter is rated by IBM at about 7.5 MIPS. How does the FLEX-ESL compare? Let's attempt to find out.

To test the FLEX-ESL processor performance, I ran several benchmark programs on a PC Server 330 System/390 with a P/390E adapter. I also ran the same benchmark programs on the FLEX-ESL system. Both systems were running VSE/ESA 2.3 in native mode.

Test                P/390E CPU          FLEX-ESL CPU         Result              

Job set 1           136 Secs            112 Secs             17 % Faster         

Job set 2           101 Secs             97 Secs             4  % Faster         

Job set 3            99 Secs             96 Secs             3  % Faster

P/390E vs. FLEX-ESL CPU Comparisons

The CPU usage of each job stream was taken from VSE/ESA job accounting data. Based on these and other benchmark job streams I found the FLEX-ES emulator to be very comparable to the IBM P/390E adapter in terms of CPU performance. In nearly every test benchmark, the FLEX-ES emulator and the P/390E adapter used the same or almost the same amount of CPU to process the same benchmark job stream.

I should point out that the CPU optimizations of the P/390E adapter and the FLEX-ES emulator are different. The FLEX-ES emulator does its work by compiling the S/390 opcodes into Intel instructions and executing the Intel instructions. Simple S/390 instructions, LR (Load Register) for example, compile into only 3 Intel instructions. Complex S/390 instructions, PC (Program Call) for example, compile into hundreds of Intel instructions. The average S/390 instruction compiles into 17 Intel instructions. What does all this mean? Your mileage will vary. The speed of the FLEX-ES emulator is about the same as the CPU on a P/390E adapter; but since the optimizations are different, your results may vary slightly from those reported here.

The VM/ESA SIE Instruction

Since the FLEX-ES emulator compiles the S/390 opcodes into Intel instructions, it makes sense that the SIE (Start Interpretive Execution) instruction used by VM/ESA would be one of the most complex. In fact, initiating the SIE environment and breaking out of the SIE environment is expensive. Benchmarking of this process shows the FLEX-ES emulator to be on par with IBM 9121/9221 processors. To minimize VM/ESA overhead look into caching data within the guest virtual machine. For VSE/ESA guest machines, Barnard Software, Inc.'s OPTI-CACHE product has been shown to reduce virtual machine CPU by as much as 30% by simply eliminating SIE and I/O overhead. This is something to keep in mind if you are planning to run VM/ESA on your FLEX-ESL system.

FLEX-ESL I/O Speed

FSI claims that the FLEX-ESL can handle much higher I/O rates than an equivalent PC Server 330 System/390 based on a P/390E adapter. Can it? Let's find out. To test this claim I ran a series of benchmark job streams on P/390E and FLEX-ESL systems running VSE/ESA 2.3 in native mode.

Test                      P/390E                     FLEX-ESL                   

Maximum I/O               700-750 / Sec              1800-1900 / Sec            

Peak Sustained I/O        350-400 / Sec              1100-1200 / Sec            

Avg. Sustained I/O        150-200 / Sec               650-750  / Sec

P/390E vs. FLEX-ESL I/O Rate Comparison

The Maximum I/O value represents the maximum number of VSE/ESA Execute Channel Program requests that the benchmark program could execute per second. On a real system this figure is not attainable because the benchmark program was executing channel programs issuing no-operations to the channel and not actually transferring data. The Peak Sustained I/O value represents the highest sustained I/O rate observed by the benchmark program while transferring data. The Average Sustained I/O value represents the average sustained I/O rate while the benchmark program was transferring data and the system was running at 100% CPU.

The factor limiting the performance of a P/390E system has always been the I/O subsystem. While the P/390E CPU is rated at 7.5 MIPS, the I/O subsystem is not fast enough to keep the processor busy. As you can see, the FLEX-ESL I/O subsystem design improves performance dramatically. The Average Sustained I/O rate of the FLEX-ESL system is about 3 times higher than the P/390E system.

DASD Observations

The FLEX-ESL emulation of CKD/ECKD devices is very good and better than the FBA emulation from a performance viewpoint. This result is the reverse of the PC Server 330 System/390 results where FBA emulation has much better performance than CKD/ECKD emulation.

The FLEX-ES DASD emulator allows you to use PC memory as cache for the FBA/CKD/ECKD devices and enhances I/O subsystem performance dramatically. Considering the fact that PC memory is very inexpensive, buy as much PC memory as possible to optimize the FLEX-ESL system's performance.

Selection of a SCSI RAID adapter and the RAID type to be used is very important. Based on test benchmarks, I would recommend using a three channel SCSI RAID adapter. Define the physical RAID arrays so that the devices are striped across the available SCSI channels and use a stripe size of 32K. Define the RAID array as RAID 0+1 (stripe and mirror) for best performance. RAID 5 arrays can also be used if you are willing to give up a little performance over the RAID 0+1 array configuration. Use of PC memory to cache the RAID 5 DASD can overcome most of the performance reduction created by using RAID 5 instead of RAID 0+1. Tuning of the RAID stripe size and SCSI channels is very important to I/O subsystem performance.

Caching Data Within VSE/ESA

High speed virtual storage DASD caching products, like OPTI-CACHE from Barnard Software, Inc., have long been used to improve system performance. Will OPTI-CACHE still improve system performance on a FLEX-ESL system?

Test            Duration         Effect          CPU              Effect          

No O/C          304 Seconds                      112 Seconds                      

O/C Run #1      175 Seconds      -42%            98 Seconds       -12%            

O/C Run #2      105 Seconds      -65%            96 Seconds       -14%

Cache Performance Using OPTI-CACHE

The first test run shows the benchmark job stream running 304 seconds using 112 seconds of CPU time (36%). The O/C Run #1 shows the effect of starting up OPTI-CACHE on the system. Since OPTI-CACHE had just initialized, its internal cache was empty. Even so, the benchmark job stream ran in 175 seconds using 98 seconds of CPU time (56%). This represents a 42% reduction in run time. The O/C Run #2 shows the effect of having all of the required data already in the cache. This benchmark run took only 105 seconds to complete using 96 seconds of CPU time (91%).

It is easy to see that caching the data within VSE/ESA still shows excellent performance gains. Run times are reduced allowing jobs to complete faster; but another effect is also shown. CPU time used by the benchmark job stream also decreased. This is because it is faster to cache an I/O request than to actually translate the channel program and start the I/O operation. In the case of this benchmark job stream, caching using OPTI-CACHE actually increases the effective MIPS rating of the FLEX-ESL system by 13%. How? CPU saved by caching is available to process additional work effectively increasing the MIPS rating. Based on this and other benchmark job streams, I found 5% to 26% reductions in CPU times when OPTI-CACHE was active. These CPU reductions effectively increased the MIPS performance of the FLEX-ESL system by up to 25%.

Cost Effective Enterprise Computing Solution

It certainly appears that FSI's claims are justified. By combining Intel PC hardware with an IBM ESL software license, FSI has created a new enterprise computing platform. Whether you want to demo mainframe software on a notebook computer or run your enterprise on an Intel server, the FLEX-ESL provides an inexpensive and very cost-effective solution to many user's computing requirements.

Reprinted by permission