 






          1  Abstract

          The turbo PrintServer 20 controller is a performance enhance-
          ment of the original PrintServer 20 system Controller. The turbo
          controller was developed to enable PostScript code to execute
          faster and thus improve page throughput for complex documents.
          The RETrACE analysis system was designed to analyze the per-
          formance of the original PrintServer 20 system and estimate
          expected performance of future systems. The turbo controller's
          processor and its three subsystems for memory, write buffer,
          and bit-map data transfer were selected based on the analy-
          sis results. Performance tests conducted on both the original
          and the turbo PrintServer 20 indicate the enhanced processing
          performance of the turbo controller. [The SERVER paper starts
          here.]

          In 1988 the turbo controller project was conceived as a means
          of extending the life of the PrintServer 20 platform by in-
          troducing a performance-enhanced system controller. The system
          controller in the PrintServer 20 is housed within and powered
          by the printer or "print engine"; it is a concise implemen-
          tation of a single-board computer containing a CPU, a memory
          subsystem, an Ethernet interface, and a printer interface. It
          supplies an environment in which a multitasking software system
          manages communications with remote clients and with the print
          engine, performs data conversion from the page description lan-
          guage (PostScript) to bit-map images, and provides management of
          physical print engine resources.

          The original controller provided a maximum print speed of 20
          pages per minute, but this performance could not be maintained
          when the document included complex text, graphics, or images. To
          improve page throughput for complex documents, a controller was
          needed on which PostScript code could execute faster. To enhance
          performance, the competition was moving toward controllers
          based on new industry-standard reduced instruction set computer
          (RISC) processors. Therefore, to be competitive, Digital's new
          controller was required to improve performance by five to eight

                                                                         1

 






          times that of the original controller, which had been based on
          the rtVAX microprocessor.

          As challenging as the performance improvement would be to
          achieve, budgetary pressures forced restrictions on the im-
          plementation strategy. We were to use existing, qualified chips
          wherever possible in order to avoid new part qualification costs
          and application-specific integrated circuit (ASIC) development
          costs.

          Early investigations indicated that the performance target was
          indeed achievable with existing inexpensive RISC processors, as
          well as a high-speed Digital proprietary VAX processor. A RISC
          processor would require porting a 2.5-megabyte (MB) software
          system, which was far beyond the scope of the project. The
          highest performance VAX processor and the associated support
          chips, which would not cause a problem with the software system,
          were far too expensive to be considered. Alternatives were
          therefore limited to less expensive, lower speed VAX processors:
          the low-risk, 60-nanosecond (ns) CMOS VAX or CVAX processor was
          proven, and the higher speed and more cost-effective "system on
          a chip" or SOC processor was under development. Either choice
          would have a minimal impact on the software system and would
          provide a cost-effective solution.

          The original performance estimates for the CVAX and the SOC
          processors in general-purpose processing environments were
          below the lower bound of the performance target. The design
          team was also uncertain of the actual execution characteristics
          of the PrintServer software. For these reasons, it was decided
          to begin the project with a performance analysis of the original
          controller to determine the expected performance of a design
          based on either processor.

          This paper discusses the problems encountered during our anal-
          ysis and the solutions devised by the Hardcopy Systems En-
          gineering Group to overcome them. The RETrACE tool suite, a
          performance analysis system, is described and analysis results

          2

 






          are provided. The paper then discusses the hardware architec-
          ture of the turbo controller and ends with a presentation of the
          performance results obtained for standard PostScript benchmarks.


          2  Performance Analysis of the Original Controller

          The PrintServer 20 software system consists of a VAXELN operat-
          ing system, an Adobe Systems, Inc. PostScript interpreter, and
          a substantial amount of software to manage communications and
          resources. The task of analyzing its performance was complicated
          by two additional factors. First, the software system's behavior
          depended on the characteristics of the user's PostScript docu-
          ment. PostScript is an interpreted programming language. Thus,
          like any computer program, low-level machine performance can be
          dramatically affected by the program being executed. Second, and
          more painful, the proprietary nature of the PostScript inter-
          preter prohibited us from obtaining code sources, and discussing
          its internal architecture with engineers from Adobe Systems.

          While the characterization of a complex, partially proprietary,
          real-time software system is difficult, it is not impossible.
          Programmer counter address (PC) traces have offered many systems
          designers very detailed insight into the execution performance
          and characteristics of systems. PC traces provide a means to
          observe a system at a macroscopic level, allowing a view of
          the complex interactions between the hardware and software
          systems. System designers can use captured address traces from
          current machine performance to extrapolate expected performance
          of future systems and help them make architectural trade-offs.

          The RETrACE Analysis System

          The RETrACE tool suite was created to provide a nonintrusive
          means of capturing real-time PC traces and analyzing the cap-
          tured addresses. The tool suite consists of both hardware and
          software components.


                                                                         3

 






          In order to keep expenses at a minimum, existing hardware was
          used wherever possible. Only one small module had to be devel-
          oped to complete the RETrACE hardware platform.

          The RETrACE hardware consists of the following:

          o  Two interconnect boards boot and operate a system controller
             on a table top. Developed as part of the original PrintServer
             20, the boards connect the controller to a print engine and
             an Ethernet.

          o  The PrintServer 20 server controller was modified for use as
             an intelligent trace buffer system.

          o  The PrintServer 20 server controller's memory capacity (12MB)
             was extended using the standard 4MB memory module used on the
             Kanji version of the PrintServer 20.

          o  The RETrACE mother board was developed specifically for this
             tool suite. It contains a 32-bit wide, first-in, first-out
             (FIFO) buffer and two loosely coupled state machines.

          o  A standard PrintServer 20 system controller and print engine
             were used as the "system under observation."

          o  The console terminal was selected from the standard VT series
             of terminals.

          A diagram of the RETrACE hardware system is shown in Figure 1.

          The RETrACE mother board performed the data capture, using the
          modified controller's memory as a large buffer. The board mon-
          itored the processor bus of the system under observation by
          copying all addresses and communications between the rtVAX pro-
          cessor and its external floating-point unit. This copied data
          was placed into a FIFO buffer that in turn was written into the
          memory of the modified controller using a direct memory access
          (DMA) device. Since a standard PrintServer 20 controller and
          its optional memory expansion provide 16MB of storage, approx-
          imately 3 seconds of real-time execution address traces could
          be captured. The data capture continued until the trace buffer

          4

 






          memory was exhausted, at which point the data was uploaded over
          a network connection to a VAX VMS computer for analysis.

          Due to the design of the original PrintServer 20 system, many
          large data areas and code sections were mapped into different
          explicit memory spaces. This subdivision provided a means of
          determining which code function was executing in any given
          segment of the address trace. With a simple statistical study it
          was possible to generate software execution histograms and to
          determine many of the characteristics of the system, including
          translation buffer, floating point, instruction stream (I-
          stream) versus data stream (D-stream), read versus write, and
          interrupt performance. Hit rates for fully associative caches
          of separate I-stream and D-stream, as well as a combined I-
          and D-stream cache, were also provided. These hit rates were
          determined for first-level write-through caches from 128 bytes
          up to 256 kilobytes (KB). Thus an upper bound for an optimum-
          performance cached memory system was determined.

          Both processors under consideration possessed the ability to
          access a memory subsystem at speeds greater than that achiev-
          able with existing low-cost dynamic random-access memory (DRAM)
          technology. The performance numbers predicted by the processor
          groups indicated that cached memory subsystems were required.
          Because these subsystems can be expensive and their performance
          is subject to the peculiarities of the software that executes on
          them, a multilevel memory simulator was developed to allow accu-
          rate studies to be performed on proposed cache architectures.

          The simulator was configured at run-time to simulate an arbi-
          trary hierarchical memory system that was N levels deep, with an
          arbitrary size, associativity, performance, and behavior at each
          level. The memory level nearest the processor was defined as
          the first level, and the last as main memory. The simulator pro-
          cessed a trace file by walking each address in the file through
          the memory hierarchy starting nearest the processor at the first
          level. If a copy of the address was found at a given memory
          level, then a hit was signaled and the next address was pro-
          cessed. If that address was not found, then a miss was signaled

                                                                         5

 






          and the simulator would proceed to the next level of memory in
          the hierarchy.

          Whenever a hit occurred at a given level, it was logged and
          all levels of memory in the hierarchy above it would allocate
          entries based on their defined allocation rules. While this
          procedure indicated the memory system performance for a proposed
          architecture, the overall system performance was still unknown.
          Using a simple rule based on the average execution time per
          address for the existing controller, and scaling that time based
          on the clock speed increase of proposed processors, an overall
          performance number was estimated for a system based on either
          processor with any arbitrary memory architecture.

          Benchmark Selection

          The RETrACE tools suite provided the components required to
          study the execution characteristics of the PrintServer system
          without changing the characteristics of its normal operation.
          The only difficulty was to narrow the focus of the benchmark
          list to provide a representative sample of PostScript documents
          to print. Due to time constraints, the list was limited to five
          benchmarks.

          BM1  The BM1 benchmark stresses those aspects of the system
          that convert the mathematical representations of characters
          to bit-map representations, which comprise the form that is
          printed. This benchmark uses several fonts in standard character
          orientations, stressing both very large and small character
          sizes.

          BM2  Of the same type as BM1, this benchmark stresses the trans-
          forms from mathematical to bit-mapped character representations;
          however, the characters printed are at arbitrary orientations
          with sizes ranging from typical to very small.




          6

 






          BM3  The BM3 benchmark is one of the standard benchmarks for
          PostScript performance qualification. It is a simple 41-page
          document that contains several different fonts. The benchmark is
          designed to characterize the standard text-handling performance
          of a printer. This benchmark is printed twice to ensure that all
          characters to be printed have been converted from mathematical
          outlines to bit-map representations of the characters. Thus
          the focus of the benchmark is to move the text data through
          the system, to copy the character bit maps to the 1MB region in
          memory that contains the image to be printed, and to print the
          image. It should be noted that this is the only benchmark that
          printed at engine speed on the PrintServer 20 system controller
          powered by the rtVAX system.

          HOUSE  A binary image file, the HOUSE benchmark was used to
          stress the communications aspects of the PrintServer system.

          SCHEM  The SCHEM benchmark was a vector representation of
          a logic schematic. This benchmark was used to stress the
          PostScript interpreter's ability to interpret nonnative
          PostScript code and to exhibit the characteristics of drawing
          vectors.

          Analysis Results

          The thrust of the analysis was to provide credible evidence to
          support architectural and implementation trade-offs. The major
          areas of focus were

             Memory system organization
             Printer interface performance
             Main memory bandwidth
             Overall system performance






                                                                         7

 






          Memory System Organization  The statistical analysis of the
          trace information provided many clues to direct our investiga-
          tion toward the optimum memory system architecture. The overall
          read-to-write ratio for the observed benchmarks ranged from as
          low as 4.3:1 up to 5.5:1, which means for a write-through cache
          system with a theoretical 100 percent read hit rate, the writes
          would degrade the overall hit rate to approximately 81 to 84
          percent. As the analysis of the data progressed, it was under-
          stood that the write data must be studied very closely since
          it could have a dramatic impact on the overall cache miss rate.
          During the cache model simulations, the hit rates of the I-
          stream were between 85 to 90 percent. However, the D-stream hit
          rates were between 35 to 45 percent, with writes accounting for
          60 to 90 percent of the total D-stream misses. To achieve the
          greatest positive effect on the hit rate of the system, enhance-
          ment of write-miss performance was the most advantageous. The
          two options to improve this performance were either to imple-
          ment a write-back cache or to add a write buffer to the system.
          Further cache simulations showed that a write buffer would pro-
          vide an 8 to 16 percent overall system performance improvement,
          which was equal to that of a write-back cache. The write buffer,
          however, was the more straightforward solution to implement.

          Cache analysis revealed that the processors required different
          memory architectures. The CVAX had an internal 1KB, two-way set
          associative cache. This was to be configured as a mixed I- and
          D-stream cache. An additional 32KB to 64KB, two-cycle write-
          through cache was to be added externally. This would also be
          configured as a mixed I- and D-stream cache. A single-longword,
          two-cycle write buffer would provide enough buffering to reduce
          the dramatic impact of write misses. The SOC was proposed to
          have an internal write-back cache between 5KB and 8KB, with each
          1KB region making up a single set. Cache simulations indicated
          that with a minimum internal mixed I- and D-stream cache of 5KB,
          five-way set associative, an external data cache would have to
          be over 64KB to have even a negligible effect on overall system
          performance. Therefore no external cache was recommended. To


          8

 






          mitigate the write-miss penalty, a two-cycle write buffer of 4
          to 6 longwords was recommended.

          As an acceleration technique, the original PrintServer 20
          controller contained a memory access capability that allowed
          data written to memory to be logically ORed with data that was
          already stored. This technique was particularly useful when
          the software system was writing the image that was ultimately
          printed. As part of the process of generating an image to print,
          the individual characters appearing on a page must be copied
          from a region of memory called the font cache to another region
          called the frame buffer. The frame buffer contains the actual
          data that is sent to the print engine. To complicate things, the
          data written to the frame buffer must be able to overlay data
          that may already be there, thus requiring a logical OR function.

          When a document was printing at or near the maximum engine speed
          of 20 pages per minute, analysis showed this low-level copy-
          ing function consumed approximately 20 percent of the total
          system time allotted to generate and print one page. Thus a
          logical OR function in the memory system would reduce the num-
          ber of memory data cycles from "2 reads 1 write" to "1 read 1
          write," and reduce the impact from a second read occupying a
          useful cache location. Without this capability, the degradation
          would be between 5 and 10 percent of overall system performance
          when printing at or near 20 pages per minute. Therefore memory
          capability with a logical OR function was recommended.

          Printer Interface Performance  When a PrintServer 20 is print-
          ing, every page that exits the printer requires the 1MB frame
          buffer to be copied from memory to the print engine interface.
          Changing a program-controlled printer interface to one driven by
          a DMA device provided two significant advantages. The first was
          to reduce the real-time requirements on the PrintServer soft-
          ware system, and the second was to allow for a limited degree
          of parallelism on the controller. The parallelism was due to
          the ability of the processor to continue to execute from its
          cache memory system while the DMA device accessed memory. The
          processor only stops executing when a cache miss occurs.

                                                                         9

 






          Main Memory Bandwidth  With a CVAX processor configured as rec-
          ommended in the section Memory System Organization, the main
          memory system bandwidth requirement of the processor was 60
          percent. For the SOC, it was 70 percent when an existing DRAM
          controller was used. A DMA-driven printer interface required 15
          percent, and an Ethernet interface required nominally 4 percent
          with bursts up to 20 percent. Each subsystem was scrutinized
          to reduce its required memory bandwidth. The resulting rec-
          ommendation was to add a 32-bit bus to the memory subsystem
          to provide a dedicated channel for all data being sent to the
          printer interface. This provision would reduce required memory
          bandwidth for the printer interface from 15 percent to about 7
          percent. The system would then have a nominal memory bandwidth
          requirement of 71 percent for a CVAX system and 81 percent for
          an SOC.

          Overall System Performance  The execution characteristics of the
          original
          PrintServer 20 provided some interesting surprises. Most
          floating-point calculations were performed in double precision;
          and even more interesting, for each floating-point operation,
          there was a floating-point conversion from single to double pre-
          cision, and then back again. Since the precise operations were
          not required, a simple compiler switch removed the conversions
          and provided a 3 percent overall system performance improve-
          ment for floating-point-intensive PostScript documents. A second
          surprise came from the results of the BM3 benchmark, which in-
          dicated a translation buffer hit rate of 85 percent. At the
          time of the discovery, the PrintServer 20 was configured with a
          standard MicroVAX processor; however, by substituting an rtVAX,
          which uses one less memory access to reference its page tables,
          an 11 percent system performance improvement was achieved. With
          this improvement, the rtVAX processor provided enough power to
          allow the original PrintServer 20 to ship with its 20-page-per-
          minute designation. This information led the turbo controller
          designers to determine that the translation buffer of the SOC
          would be large enough for all the entries required.


          10

 






          Results

          The final analysis revealed that the expected performance of a
          CVAX or SOC processor would place either design on the low side
          of the performance requirement. Therefore close attention to
          detail would be required during the implementation phase of the
          project as every ounce of performance mattered. The expectation
          was to have a choice between an SOC processor with a 40-ns
          cycle time and a CVAX processor with a 60-ns cycle time. The
          performance improvements of the two processors are compared in
          Table 1.

          As the project schedule progressed, the risk associated with the
          new SOC processor decreased. As this risk window collapsed,
          it was understood that a turbo controller based on the SOC
          processor would not only perform better, but would also cost
          less as it would not require an external cache.

          ________________________________________________________________

          Table 1:  Performance Improvement Relative to Original PrintServer
          __________20_Controller_________________________________________

          Benchmark   SOC         CVAX
                      Procesor    Processor

          BM1         4.7         3.7

          BM2         4.9         4.0

          BM3         4.3         3.3

          HOUSE       4.9         4.2

          SCHEM_______4.7_________3.7_____________________________________




                                                                        11

 






          3  Turbo Controller Hardware Design

          The turbo controller was destined for a relatively high-end
          printer. Therefore the hardware architecture had to provide
          maximum performance, even though this implementation would
          increase costs. Based on the results obtained during RETrACE
          analysis, the hardware design had the following implementation
          goals:

          o  The SOC would provide the CPU, the floating-point accelerator
             (FPA), and the cache subsystem. No second-level cache would
             be implemented.

          o  A four- to six-entry write buffer would be implemented.

          o  The transfer of bit-map data to the print engine would re-
             quire a 32-bit DMA subsystem with scan-erase capability.

          o  The memory subsystem would support OR-mode memory access by
             the CPU and scan-erase access by the DMA controller.

          Although both the SOC and rtVAX chips comply with the VAX archi-
          tecture standard and both are conceptually very similar, they
          have significant differences in the bus interface. For example,
          the SOC uses a quadword cycle (one 32-bit address followed by
          two 32-bit data reads) to fill one internal cache block, while
          the rtVAX processor, which does not support caching, does not
          use this type of cycle. Also, the clocking system on the SOC
          was enhanced, and the timing relationships between signals were
          modified to improve performance.

          The changes to the SOC bus interface, plus the required func-
          tional changes revealed by RETrACE analysis, meant that very
          little of the original PrintServer 20 controller design could be
          applied to the new controller. One of the first questions to be
          answered before the design of the turbo controller could begin,
          was whether or not one or more ASICs would be required for the
          design. This question had to be answered for three subsystems:

             Main memory
             Write buffer

          12

 






             Bit-map data transfer subsystem

          In each case existing chips satisfied some of the requirements
          for the subsystem. In the end these chips met all our require-
          ments, but only because they were used in ways not originally
          intended by the chip designers.

          Main Memory

          Since the SOC has a bus interface that is compatible with the
          CVAX chip, the most obvious chip to use as a main memory con-
          troller was the CVAX memory controller (CMCTL) chip.[1] It re-
          sponds to all bus cycles generated by the SOC, and since it was
          already used on a number of platforms supported by the VAXELN
          operating system, its use greatly simplified porting VAXELN to
          the turbo controller. However, the turbo controller requires
          two special memory modes that are not provided directly by the
          CMCTL, namely OR mode and scan-erase mode. It was essential to
          devise a way to include these two modes if the CMCTL were to be
          used.

          OR-mode memory is a technique used to improve performance during
          the writing of the page bit map into memory (scan conversion).
          During normal memory operation (called replace mode), the desti-
          nation operand in memory is replaced by the source operand. Dur-
          ing an OR-mode write cycle, the destination operand is modified
          as follows:

          o  For each logical zero in the source data being written, the
             corresponding destination bit in memory remains unchanged.

          o  For each logical one in the source data being written, the
             corresponding destination bit in memory is written with the
             corresponding bit in the pattern register.

          o  The pattern register is a 32-bit register which determines
             the "color" pattern of the "ink" being written on the page.


                                                                        13

 






          Figure 2 shows a portion of the logic between the CMCTL and the
          memory array that implements the OR-mode function in hardware.
          The OR-mode operation is accomplished by inverting the source
          data and connecting it to 32 independent write enables of the
          memory array. When a zero is written, it is inverted and the
          write cycle for that bit becomes a read cycle, thus preventing
          any change to the memory contents. When a one is written, it
          is inverted and the write is allowed to occur, but the data
          actually written depends on the value previously written into
          the pattern register.

          Two features of the CMCTL chip make it possible to implement
          OR-mode memory. First, its 64MB address space is divided into 4
          arrays of 4 banks (16 banks total). Second, the CMCTL chip can
          selectively disable parity checking on an array.

          The large address space of the CMCTL allows the use of 2 arrays
          for replace mode and 2 arrays for OR mode, since the turbo con-
          troller supports up to 32MB of memory. The control signals of
          the two sets of arrays are combined such that OR mode and re-
          place mode access the same physical memory, though in different
          ways. Parity error detection is disabled on the OR-mode arrays;
          thus a read-through OR-mode address space cannot cause a parity
          error. This is necessary because OR-mode write cycles may cor-
          rupt parity. Normally any bit map created using OR-mode write
          cycles is read using OR-mode read cycles.

          The other special mode required for the main memory system
          is called scan-erase mode. It is an operating mode designed
          to improve bus utilization during the transfer of the bit map
          from main memory to a FIFO buffer connected to the printer data
          lines. This mode is made possible by a side effect of the error-
          correcting code (ECC)/parity generation logic in the CMCTL.
          Any time a masked write occurs (any write other than an aligned
          longword, such as a byte write), the destination longword must
          first be read by the CMCTL, then combined with the bytes to be
          written in order to generate the parity or ECC check bits for
          that longword.

          14

 






          Three operations occur during a single scan-erase cycle. Refer
          to the circuit drawing in Figure 3.

          1. The bus master asserts the signal's "bit-map load" and "bit-
             map erase" and requests a masked write. The CMCTL performs a
             read, and the bit map is read onto the memory data bus.

          2. Bit-map data is automatically transferred from the memory
             data bus into the FIFO buffer.

          3. The CMCTL performs a write. However, since the bit-map erase
             signal has disabled the data path and the pull-down resistors
             have set the data-in lines to all zeros, the write cycle,
             which was intended by the designers of the chip as a masked
             write, has in fact become a memory clear operation.

          Write Buffer

          The LR3220 chip was chosen as the base for the write buffer sub-
          system. It provides a six-entry FIFO buffer for address, data,
          and byte mask and detects whether the processor has requested a
          read at a memory location for which a write is still pending. It
          also supports two operating modes: LR3000 mode and Harvard mode.

          If it were not for the Harvard-mode feature, it would have been
          more difficult to include the LR3220 chip into the turbo con-
          troller. The LR3000 processor, for which this chip was designed,
          has staggered address timing. Some of the address and byte-mask
          bits are asserted on the falling edge of the clock, and the re-
          maining bits are asserted on the rising edge of the clock. When
          the LR3220 chip is configured in LR3000 mode, the processor sub-
          system must meet these timing requirements. However, when the
          LR3220 chip is configured in Harvard mode, all address, data,
          and byte-mask information is read at the same rising clock edge.

          The basic strategy for including the write buffer into the
          turbo controller was to insert the write buffer between the
          SOC and the rest of the system as shown in Figure 4. The SOC
          would issue read and write requests to the write buffer, and the
          write buffer would issue read and write requests to the rest of

                                                                        15

 






          the system. During CPU cycles the SOC and the write buffer have
          a master-slave relationship in which the SOC is the master. The
          relationship between the write buffer and the rest of the system
          is also a master-slave relationship; however, the write buffer
          is the master. In fact, the write-buffer output interface must
          look almost identical to the SOC.

































          16

 






          The structure of the write-buffer subsystem is shown in Figure
          5. The bus interface unit responds to read or write requests
          from the SOC. During write cycles, the bus interface writes
          the data into the LR3220 chip and immediately alerts the SOC
          to terminate the cycle quickly. Whenever one or more entries in
          the LR3220 chip have data, the bus cycle generator (BCG) removes
          the next entry and issues a write request to the appropriate
          subsystem.

          The write-buffer subsystem allows the SOC to "read around" the
          write buffer, provided the address being read does not have a
          pending write in the LR3220. To handle this, the BCG includes
          an arbitration circuit. When the SOC requests a read cycle,
          the bus interface unit of the write buffer passes the request
          to the BCG. The BCG responds once it has completed any write
          cycle currently in progress, provided that the address to be
          read does not have a pending write in the write buffer. When the
          slave device being read acknowledges the BCG, the acknowledgment
          is passed back to the bus interface and finally to the SOC to
          terminate the cycle. The BCG then resumes its task of removing
          entries from the LR3220 chip and issuing writes to the rest of
          the system.

          In order to maintain data coherency, the write-buffer subsystem
          enforces some additional protocols.

          o  All writes to any location other than main memory require
             a write-flush cycle; that is, the bus interface must wait
             until the LR3220 chip is empty before writing the data to it.
             Furthermore, the bus interface must wait until the BCG has
             finished the cycle before it acknowledges the SOC and allows
             it to perform the next cycle.

          o  All reads to any location other than main memory require a
             read flush, which has the same restrictions as a write flush.
             These restrictions are required to avoid the possibility of
             reading around a pending I/O space write, which often has
             side effects to other addresses.

                                                                        17

 






          o  The write-buffer subsystem must pass all DMA bus transactions
             to the SOC to ensure that all cached memory locations that
             are modified by DMA cycles have their corresponding cache
             entry invalidated.

          Bit-map Transfer Subsystem

          The bit-map transfer subsystem transfers bit-map data, cre-
          ated by the PostScript interpreter, to the print engine. It is
          composed of the 32-bit DMA controller, a FIFO subsystem, and
          scan-erase logic in main memory as described in the section Main
          Memory.

          The main requirements for the 32-bit DMA controller were

          o  32MB address range

          o  Ability to transfer 32 bits at a time

          o  Ability to transfer the frame buffer forward (incrementing
             the source address) or backward (decrementing the source
             address)

          None of the available DMA controller chips met all our require-
          ments, but the AMD 9516 universal DMA controller (UDC) met some
          of them. The UDC is a 16-bit DMA controller with a 16MB address
          range and the ability to increment or decrement the source ad-
          dress. There were two drawbacks to the use of this chip. The
          software would have to ensure that the frame buffer was always
          within the lower 16MB of memory, and the UDC would use twice
          as much bus bandwidth since it could transfer only 16 bits at a
          time.

          It was proposed that the UDC could be used as a full 32-bit
          DMA controller if it was connected to the bus "incorrectly" by
          shifting the data/address lines to the left by one bit. That
          is, data/address line 0 on the UDC would be connected to data
          /address line 1 on the bus; data/address line 1 of the UDC would
          be connected to data/address line 2 on the bus; etc. This type
          of connection doubles the address range of the chip and causes

          18

 






          the source address on the bus to increment by 4 bytes (32 bits)
          instead of 2 bytes (16 bits).

          This decision had a few implementation impacts. For example,
          the register definitions were now incorrect, since all the
          bits in all the registers were shifted one bit to the left.
          However, once the software was modified to compensate for this,
          the UDC functioned properly as a 32-bit DMA controller. When
          combined with the scan-erase feature of main memory, it allowed
          us to achieve our bit-map transfer goal of reading 32 bits from
          memory, loading it into the FIFO subsystem, and clearing the
          memory location, all in a single DMA cycle.



























                                                                        19

 






          4  Performance

          In this section, the performance of the original PrintServer 20
          is compared to the enhanced performance of the turbo PrintServer
          20.

          Except for performance, the original PrintServer 20 and the
          turbo PrintServer 20 have identical functional capabilities.
          Table 2 lists the five functional subsets that were character-
          ized for performance on both printers. The first four functional
          subsets were rated using the PostScript real-time operator; they
          measure the elapsed CPU time needed to complete a test. The last
          functional subset was rated according to the rate of pages ex-
          iting the printer. The term "DECnet/DPS" refers to the DECnet
          job (a job is one of several multiprocessing tasks running on
          the controller) and the "distributed PrintServer software" job.
          The term "printer system" refers to the complete printer system,
          including the PostScript job and the printing overhead jobs. The
          printer system was rated according to the rate of pages exiting
          the printer.

          ________________________________________________________________

          Table_2:__Functional_Subsets_of_the_Printers____________________

          Functional         Characterization
          Subsets

          PostScript job     Math operations per second

          PostScript job     Text: characters per second

          PostScript job     Graphics: vector inches per second

          DECnet/DPS jobs    DECnet/DPS: kilobytes per second

          Printer system     Image printing: square inches per second
          excluding
          DECnet/DP_______________________________________________________

          20

 






          Table 3 reports the general attributes of the five files that
          were run with the RETrACE system and characterized for perfor-
          mance.

          ________________________________________________________________

          Table_3:__Benchmark_File_Attributes_____________________________

          File        General Attributes of File
          Name

          BM1.PS      Contains 39 pages of text with 13 fonts of various
                      sizes. Some text strings are at varying angles.

          BM2.PS      Contains 1 page of spiral text of various point
                      sizes.

          BM3.PS      Contains 41 pages of text with 5 fonts.

          HOUSE.PS    Contains a 1-page bitonal image of 3000 blocks
                      (DECnet limited).

          SCHEM.PS    Contains a 65-page schematic of graphics (vectors)
          ____________and_text.___________________________________________

          Math Operators Performance of the PostScript Job

          Figure 6 illustrates the controllers' performance results in
          math operations per second. The test determines the time needed
          to perform 50,000 primitive math operators (e.g., adding two
          numbers 50,000 times) during a PostScript test document. The
          real-time operator reads the current time, and the repeat
          construct repeats the math operator. This test measures the
          performance of the CPU only.





                                                                        21

 






          Text Performance of the PostScript Job

          Figure 7 compares the text performance of the PostScript job
          on the original controller and the turbo controller. The test
          determines how long it takes the PostScript job to compose
          250,000 equally sized characters to the page buffer in memory,
          which eventually is sent to the print engine to be printed.

          Graphics Performance of PostScript Job

          An important means of characterizing graphics performance is in
          vector inches per second. Figure 8 shows the results obtained by
          running a PostScript vector program in which all vectors are at
          45 degrees and vector lengths are from 0.1 inch to 3 inches.

          Image Performance

          The image test characterized the complete printer system, in-
          cluding the PostScript job and the printing overhead jobs, but
          excluding the DECnet/DPS time required to transfer an image file
          to a printer. Three one-square-inch bitonal images at device
          resolution were placed into the user dictionary and were used
          repeatedly during the performance measurement. The result of
          using these precached images was to eliminate the DECnet and
          DPS software time that would be required to transfer a full-page
          image from a host to the printer. Performance was measured by
          printing 10 pages of 80 square inches of image per page.

          The pages were printed landscape and portrait to measure the
          image performance both on axis and off axis. (On axis means
          that the printer sequentially prints all bits of a word from the
          image on a single scan line. Off axis by 90 degrees means that
          the printer prints one bit from each word and does not print
          the next bit in the word until it is at the same position on
          the next scan line.) Figure 9 shows the results of the image
          performance test in square inches per second.



          22

 






          DECnet/DPS Jobs Performance

          DECnet/DPS transfer rates can be ignored for text and graphics
          files, but these rates can consume most of the time needed to
          print large image files. For example, a single, letter-size page
          of image contains more than 1MB of image data, but the corre-
          sponding PostScript file contains more than 2MB. Because the
          image data is represented in American standard code for informa-
          tion interchange (ASCII) hexadecimal characters in PostScript,
          8 bits of the PostScript file are needed to represent 4 bits of
          image data.

          To measure DECnet/DPS, a PostScript file of 1MB of comments was
          sent to the printer. The clock was started when the beginning
          of the file was received by the PostScript interpreter and
          stopped when the end of the file was received. The assumption of
          this test method was that the PostScript interpreter can parse
          comment lines much faster than DECnet/DPS can transfer them.

          The DECnet/DPS transfer rate is basically proportional to the
          slower of the host and printer processors. Figure 10 shows the
          DECnet/DPS results.

          RETrACE Benchmark Files

          The benchmark files listed in Table 4 are characterized both by
          the elapsed time from file arrival to file printed and by the
          amount of CPU time used to print the job. For example, in the
          BM3 benchmark, the speed is limited by the 20-page-per-minute
          print engine, but the CPU time needed to print the file can be
          used as a performance measurement.








                                                                        23

 






          ________________________________________________________________

          Table 4:  Benchmark Files Characterized by Elapsed Time and CPU
          __________Time_(Seconds)________________________________________

          Benchmark   Original   Turbo      Original   Turbo      delta      delta
                                 CPU                   Elapsed    CPU        Elapsed
          File        CPU                   Elapsed

          BM1         7585       1707       7735       2050       4.4        3.8

          BM2         238        51         241        51         4.7        4.7

          BM3         56         15         128        120        3.7        1.1*

          HOUSE       67         15         106        31         4.5        3.4

          SCHEM_______2802_______625________3073_______675________4.5_____   4.6

          * Limited by engine.

          5  Summary

          The turbo controller enhanced the performance of the PrintServer
          20 printer system. Its design was prompted by the need to main-
          tain print speed performance for complex documents containing
          text, graphics, and images. The RETrACE system was designed
          to analyze the PrintServer 20 system to determine which ar-
          chitectural changes would provide the greatest improvement in
          PostScript performance. By optimizing hardware only in areas
          where it was truly worthwhile, we were able to use existing
          chips and reduce development costs. The subsystems of the turbo
          controller hardware that were optimized as a result of this
          analysis were the processor (SOC which provided CPU, floating-
          point accelerator, and cache subsystem), a memory subsystem with
          OR-mode and scan-erase access, a write-buffer subsystem, and a
          32-bit DMA subsystem. Results of the performance tests for five


          24

 






          benchmarks, including PostScript jobs, indicate the levels of
          enhanced performance.


          6  Acknowledgements

          Chris Mayer designed and implemented the RETrACE multilevel
          cache simulator. He developed a ticketing algorithm that simpli-
          fied the management of delayed behavior memory constructs such
          as write buffers.

          7  Reference

          1. D.K. Morgan, "The CVAX CMCTL-A CMOS Memory Controller Chip,"
             Digital Technical Journal, vol. 1, no. 7 (August 1988):139-
             143.

          Guido R. Simone Guido Simone is a principal engineer in the
          Print Systems Engineering Group and was the project leader
          and architect for the turbo PrintServer 20 controller. He is
          currently working on the development of a new print system
          architecture to be used with advanced printing technologies.
          In previous work, Guido was the project leader and architect
          for an rtVAX 78R32 CPU chip-based laser printer controller.
          Before joining Digital in 1980, he received a B.S. in electrical
          engineering from Rensselaer Polytechnic Institute.

          Jeffrey A. Metzger Presently a senior engineer, Jeff came to
          Digital as a co-op student in 1983, working first in the Semi-
          conductor Engineering Group and then in Hardcopy Engineering.
          He became a full-time employee after graduating from Cornell
          University in 1985. He introduced Hardcopy to system-level logic
          simulation, contributed to the hardware, software, and firmware
          development of the PrintServer 20, and developed RETrACE, which
          is used to characterize the execution behavior of PrintServer
          systems. Jeff is currently working in the Entry Systems Business
          Group on a next-generation processor product.


                                                                        25

 






          Gary P. Vaillette Senior hardware engineer Gary Vaillette has
          been involved in the design and implementation of printing
          system hardware since joining Digital in 1983. His current work
          includes performance characterization of PostScript printers
          and PrintServer products, and hardware implementation of CCITT
          decompression in the turbo PrintServer 20 product. Previously,
          Gary worked at Data General Corporation and helped to develop
          their token bus network product. He holds an A.A. (1974) in
          electrical engineering from Quinsigamond Community College and
          expects to receive a B.S.C.S. (May 1992) from Boston University.

          Digital, MicroVAX, rtVAX, VAX, and VMS are trademarks of Digital
          Equipment Corporation.

          PostScript is a registered trademark of Adobe Systems, Inc.
























          26
=============================================================================
Copyright 1991 Digital Equipment Corporation.  Forwarding and copying of this 
article is permitted for personal and educational purposes without fee 
provided that Digital Equipment Corporation's copyright is retained with the 
article and that the content is not modified. This article is not to be 
distributed for commercial advantage. Abstracting with credit of Digital 
Equipment Corporation's authorship is permitted.  All rights reserved.
=============================================================================
