Apparatus and method for dynamic frequency adjustment in a frequency synthesizer

[0002] The present invention is related to the following applications entitled “AN APPARATUS AND METHOD FOR HIGH RESOLUTION FREQUENCY ADJUSTMENT IN A MULTISTAGE, FREQUENCY SYNTHESIZER”, U.S. application Ser. No. 09/631,718, Attorney Docket Number AUS9-2000-0241-US1; and “AN APPARATUS AND METHOD FOR SYNCHRONIZING NODES IN A HETEROGENEOUS COMPUTER SYSTEM”, U.S. application Ser. No. 09/631,712, Attorney Docket Number AUS9-2000-0243-US1; which are incorporated herein by reference.

[0003] 1. Technical Field

[0004] The present invention relates generally to an improved method for frequency generation and in particular to an apparatus and a method for adjusting the generated frequency. Still more particularly, the present invention provides an apparatus and a method for making rapid frequency adjustments by using a frequency divider with a variable divisor value.

[0005] 2. Description of the Related Art

[0006] A phase locked loop (PLL) is a very interesting integrated circuit that blends analog and digital techniques. Although the basic design of a PLL has been known for decades, the circuit only became a practical building block in integrated circuit form where the cost has become affordable and the design has become more reliable.

[0007] The PLL contains a phase detector, an amplifier, a voltage controlled oscillator (VCO), and a feedback loop that allows the output frequency to be a replication of the input signal with noise removed or a multiple of the frequency of the input signal. PLLs have been used for demodulation of FM signals, for tone decoding, for frequency generation, for generation of “clean” signals, and for pulse synchronization, to name but a few of the applications. Because the output frequency is a multiple of the input frequency, it is difficult to make fine frequency adjustments using such a frequency synthesizer.

[0008] A non-uniform memory access (NUMA) computer system is a multiple processor architecture where there is a single memory address space but where memory is separated into “close” banks of memory and “distant” banks of memory. Access is “non-uniform” because the access times for the close banks of memory directly associated with the node that contains the CPU are much faster than the access times for distant memory banks at other nodes in the system. A distinct advantage of a NUMA architecture is that it scales well, in the sense that adding more nodes and processors to the system does not create bottlenecks that degrade performance in the same way as other parallel architectures.

[0009] One problem with NUMA architectures is to keep the nodes synchronized. Transactions are often labeled with time stamps that are generated by the time of day at-each node in the system. Since these nodes have independent clocks, even though they are initialized at precisely the same time, they will eventually drift apart and require re-synchronization. It is important to have precise time stamps with as little “cycle slippage” as possible between the nodes.

[0010] Therefore, it would be advantageous to have a frequency synthesizer that can be changed in small increments dynamically.

[0011] An apparatus and a method is presented for making small frequency adjustments in a frequency synthesizer. The frequency synthesizer consists of the forward portion of a phase locked loop with feedback through a fixed frequency divider and the output of the forward portion of the phase locked loop connected to a dynamically variable frequency divider. By changing the constant of division in the variable frequency divider, the output of the frequency divider can be rapidly changed in small increments.

[0012] The dynamically variable frequency divider is key to this design. This digital circuit stores the current divisor value and has an input for a new divisor value. When a signal is sent to switch to the new divisor value, the circuit uses an incrementer and associated logic to rapidly change to the new constant of division.

[0013] The novel features believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as a preferred mode of use, further objectives and advantages thereof, will best be understood by reference to the following detailed description of an illustrative embodiment when read in conjunction with the accompanying drawings, wherein:

[0014] FIG. 1 is a pictorial representation of a multiprocessor system with a non-uniform memory access (NUMA) architecture in which the present invention may be implemented;

[0015] FIG. 2 illustrates multiple nodes in a NUMA architecture in which the present invention may be implemented;

[0016] FIG. 3 shows a conventional frequency synthesizer as it exists in the prior art;

[0017] FIG. 4 shows a plot of instantaneous frequency error versus instantaneous phase error for a conventional frequency synthesizer;

[0018] FIG. 5 shows a frequency synthesizer as it exists in a preferred embodiment of this invention;

[0019] FIG. 6 shows a three stage frequency adjuster in accordance with a preferred embodiment of this invention;

[0020] FIG. 7 shows the detailed circuitry of a dynamic frequency divider as it exists in a preferred embodiment of this invention; and

[0021] FIG. 8 is a plot for phase and frequency adjustments as a function of time in accordance with a preferred embodiment of this invention.

[0022] With reference now to the figures, and in particular with reference to FIG. 1, a pictorial representation of a multiprocessor system with a non-uniform memory access (NUMA) architecture. This simplified diagram shows a system with three nodes: 100, 120, and 140. Node 100 contains four CPUs, 102, 104, 106, and 108, local memory 110, input/output channels 112, and communications channels, 114. Node 120 contains four CPUs, 122, 124, 126, and 128, local memory 130, input/output channels 132, and communications channels, 134. Node 140 contains four CPUS, 142, 144, 146, and 148, local memory 150, input/output channels 152, and communications channels, 154.

[0023] External disk drive 156 is connect to input/output channel 152. The nodes are interconnected using high speed channels 116 and 136. This system contains a single address space composed of memory banks 110, 130, and 150. Access of a CPU to its local memory bank, such as CPU 102 accessing memory 110, will be very fast since it does not need to use the node interconnections 116 or 136. Access by a CPU to a distant memory bank, such as CPU 102 accessing memory 130, will be slower since data must be transferred on communications channel 116.

[0024] Those of ordinary skill in the art will appreciate that the hardware depicted in FIG. 1 may vary. For example, the nodes may be arranged in a hypercube network where the number of nodes is a power of two (2_n) and each node is connected to exactly n other nodes. For example, each node in a hypercube with 16 nodes would be connected to four other nodes. The depicted example in FIG. 1 is less complex than this and is not meant to imply architectural limitations with respect to the present invention.

[0025] FIG. 1 is intended as an example and not as an architectural limitation for the processes of the present invention. The type of processor in a NUMA system may be homogeneous, but the present invention is also applicable to a heterogeneous system where processors or computers of different types are all part of multiprocessor computing environment. It is assumed these separate processors do not share a common clock so there relative frequencies can shift, as illustrated in FIG. 2 given next.

[0026] FIG. 2 illustrates multiple nodes in a NUMA architecture, such as the architecture shown in FIG. 1. The n nodes are labeled Node 0, Node 1, Node 2, and so forth through Node n−2, and Node n−1. Each node has an independent frequency source denoted f_othrough f_n−1. These frequency sources are used to generate the timestamp stored in the time of day (TOD) register. Since these nodes have independent clocks, even when the timestamps are initialized at precisely the same time, they will eventually drift apart and require re-synchronization.

[0027] Even if the nodes are architecturally similar, one of the nodes needs to be designated the master, in this diagram Node 0, and the other nodes are “slaves”, in the sense that their time of day is re-synchronized to the “master” time of day. It is important to have precise frequency generation with as little “cycle slippage” as possible between the nodes. What is required is a frequency generation system with the possibility of making fine adjustments to the system clock frequency on a dynamic basis so that the time of day register value can be changed.

[0028] FIG. 3 shows a conventional frequency synthesizer as it exists in the prior art. The input to the phase locked loop is reference frequency 302. Frequency divider 304 divides the reference frequency by K₂, which is an integer value in the range of 1, 2, . . . , N₂. The output of frequency divider 304 is fed into phase detector 306. The other input to the phase detector will be discussed below. The output of phase detector 306 is fed into charge pump 308. The charge pump creates a current for the period of time during which the phase error exists, which is integrated by capacitor C1310 to create a voltage Vc which is fed into voltage controlled oscillator (VCO) 312. VCO output is the frequency output from the circuit and equals (K₁/K₂) f_ref. This signal is fed into frequency divider 316 that divides f_outby K₁, which is an integer value in the range of 1, 2, . . . , N₁. The output of frequency divider 316 equals f_ref/K₂at steady-state and this is the second input to phase detector 306. This completes the feedback loop. Since both inputs to phase detector 306 equal f_ref/K₂, any shift in one of these frequencies will be detected by phase detector 306 and feed through charge pump 308 to voltage controlled oscillator 312. This results in f_outbeing adjusted to bring it back into sync to a value (K₁/K₂) f_ref.

[0029] The values of K₁and K₂must be fixed to avoid cycle-slipping due to PLL pullout frequency. The value of f_outis equal to (K₁/K₂) f_ref. By setting K₁and K₂to different integer values, the output frequency is synthesized based on the input frequency. However, these values cannot be changed dynamically, as explained below.

[0030] FIG. 4 shows a plot of instantaneous frequency error versus instantaneous phase error. A pure frequency error introduced from the steady-state condition at the origin will cause cycle slippage if it exceeds the pullout frequency of approximately 2.5 MHz. In FIG. 4 it can be seen that frequency deviations less than the pullout frequency will naturally go back towards the origin, however, frequency deviations greater than the pullout frequency will result in large instantaneous frequency errors and cycle slipping when the divider settings K₁and/or K₂, as shown FIG. 3, are changed. Therefore, when using a conventional phase locked loop, the values of K₁and K₂are fixed which makes fine adjustments of the output frequency impractical.

[0031] FIG. 5 shows a frequency synthesizer as it exists in an embodiment of this invention. As in FIG. 3 the input is still f_refand the output is (K₁/K₂) f_ref. The major difference is that the divide by K₂frequency divider is moved from the input of the circuit to the output of the circuit. As is detailed in the explanation below, this allows the value of K₂to be varied which results in the ability to make fine adjustments to the output frequency. Such adjustments are not possible using the circuit in FIG. 3 due to the problem of cycle slipping.

[0032] The input to the phase locked loop is reference frequency 502, which is fed into phase detector 504. The other input to the phase detector will be discussed below. The output of phase detector 504 is fed into charge pump 506. The charge pump creates a current for the period of time during which the phase error exists, which is integrated by capacitor C1310 to create a voltage Vc which is fed into voltage controlled oscillator (VCO) 512. VCO output equals K₁f_ref. This signal is fed into frequency divider 516 that divides it by K₁, which is an integer value in the range of 1, 2, . . . , N₁. The output of frequency divider 516 equals f_refand this is the second input to phase detector 504. This completes the feedback loop. Since both inputs to phase detector 506 equal f_refany shift in one of these frequencies will be detected by phase detector 504 and fed through charge pump 506 to voltage controlled oscillator 512.

[0033] Circuit output f_out514 is generated by feeding the output of VCO 510 into frequency divider 512 which divides its input by K₂to produce the value (K₁/K₂) f_ref. This is the same output value as the circuit in FIG. 3, but there is one major difference. The value of K₂can be varied without causing cycle slipping.

[0034] Of particular interest is the case where K₁is approximately equal to K₂so that the ratio K₁/K₂is equal to 1 plus or minus a small delta factor. Substituting these values in the equation for the output frequency results in f_out=(1±Δ) f_ref. So by varying the value of K₂, which can be changed without cycle slipping, the output frequency can be adjusted up or down by small amounts relative to the input frequency.

[0035] The circuit in FIG. 5 can be cascaded to produce even finer frequency adjustments. A three stage frequency adjuster is shown in FIG. 6. The front end of stage 1 is a conventional frequency synthesizer, similar to that shown in FIG. 3. In particular, frequency divider 602 divides the input frequency, f_ref, by L. This is fed to the forward path of phase lock loop 604, which contains a phase detector, such as 306, a charge pump, such as 308, a capacitor, such as 310, and a voltage controlled oscillator, such as 312. In FIG. 6 these components are lumped together and referred to as a phase lock loop.

[0036] To close the loop, the output of phase lock loop 604 is fed through frequency divider 606 where the division is by K₁. This output is fed back as the second input to the phase detector that is part of phase lock loop 604. The frequency output from this conventional frequency synthesizer is K₁f_ref/L, where both K₁and L are fixed.

[0037] To allow for dynamic frequency adjustment, the output of phase lock loop 604 is the input to frequency divider 608 that divides its input frequency by K2. The value for K2 can be varied dynamically, in a manner similar to the dynamic adjustments to frequency divider 512 in FIG. 5. The detailed circuitry of this dynamic frequency divider are disclosed in FIG. 7. The output from stage 1 of the three stage frequency adjuster is (K₁f_ref)/(K₂L) where K₁and L are fixed and K₂is variable. We label this output as f₂.

[0038] Stage 2 of the three stage frequency adjuster contains the forward path of phase locked loop 610, the feedback circuit with frequency divider 612 that divides by K₃, and the frequency divider 614 at the output that divides by K₄. The frequency output of stage 2 equals (K₃/K₄) f₂; this frequency is referred to as f₃. The value of K₃is fixed but the value of K₄is variable.

[0039] Stage 3 of the three stage frequency adjuster has the same structure as stage 2. It contains the forward path of phase locked loop 618, feedback loop with frequency divider 620 that divides by M, and frequency divider 622 on the output that divides by N. The frequency output, f_out, of this final stage equals (M/N) f₃. The values of M and N are both fixed.

[0040] Substituting the various formulas for each stage of the circuit, it can be seen that f_out=(K₁/K₂) (K₃/K₄) (M/N) (f_ref/L) where K₂and K₄are variable. It is instructive to substitute typical frequency values to see how the output frequency can be tuned with high refinement. Let f_refbe 150 MHz. The values of the various dividers will be chosen so that the output frequency will also be 150 MHz, but by varying the values of K₂and K₄, fine adjustments can be obtained. L, K₁, and K₃are set to 100. M and N are set to 200 and 2, respectively. In case 1, K₂is set to 119 and K₄is set to 84. The resultant output frequency is 150.06 MHz; this is a change of +60,000 Hz for 150 MHz or +400 parts per million (PPM). In case 2, K₂is set to 122 and K₄is set to 82. The resultant output frequency is 149.94 MHz; this is a change of −60,000 Hz for 150 MHz or −400 PPM.

[0041] By carrying through the calculations stage-by-stage, it is found the frequency shifts at stage 2 are less than 2.5% and at stage 3 are less than 0.08%. As one of ordinary skill in the art will appreciate, greater refinement of frequency adjustment can be obtained when more stages are cascaded. The frequency divider at the output of each stage, except for the last stage, can be made variable.

[0042] Frequency dividers in the prior art are hardwired to a particular divisor value. Therefore, a new circuit had to be devised that could divide by any integer value and that could change the divisor value very quickly.

[0043] With reference now to FIG. 7, a circuit diagram is given for a dynamic frequency divider. REG_A 704 holds the divisor, call it K. REG_B 712 holds the current counter value. REG_OUT 716 is a one-bit register that holds the output state. All three registers are clocked synchronously by the input frequency. MUX (multiplexer) 702 has two inputs, NEW_K and CURRENT_K, and one of these is selected based on the value of CHANGE_K. When CHANGE_K is activated, the output of the MUX is the input NEW_K. If CHANGE_K is not activated, then the output of the MUX is input CURRENT_K.

[0044] Comparator A>B COMP 706 is on whenever the current counter value is less than the current divisor value. Whenever the comparator 706 is on, the incrementer INC 710 increases the counter value by 1 and saves the new value in REG_B 712. The output state based on the setting of REG_OUT 716 remains the same. When the counter value exceeds the divisor value, then the output of comparator 706 is off, which causes the incrementer to be set back to 1 and the value of REG_OUT 716 to be toggled resulting in the output frequency changing state.

[0045] Examination of particular frequency values helps understand operation of this circuit. Suppose the output of the multiplexer is a divisor value of 120 and the value in REG_B 712 has just been reset, so REG_B 712 counts from the value 1 up to the divisor value. When this counter equals the divisor value, it triggers the output of A>B COMP 706 to change state. This has two effects: it resets the value in REG_B to 1 and it toggles the output frequency from REG_OUT 716. For every 120 pulses on the input, there is one pulse on the output. So the circuit functions like a “divide by 120” circuit.

[0046] Suppose that the value of NEW_K is 110 and the CHANGE_K command is received; this transfers the value of 110 to the “A INPUT” of the multiplexer. There are two possible cases: the counter value in REG_B is less than 110 or the counter value is between 110 and 120. If the case the counter is less than 110, REG_B 712 continues to count but now will be reset when 110 is reached. If the value in REG_B is already greater than 110, then the output of comparator A>B COMP is switched which results in toggle of the output frequency and a reset of the counter.

[0047] As one of ordinary skill in the art will appreciate, the case where NEW_K is larger than CURRENT_K is even easier. The current counter value is less than NEW_K, so once the multiplexer switches the input to the comparator, the counter will continue to count up until the new divisor value is reached.

[0048] With reference now to FIG. 8, a plot is given for phase and frequency adjustments as a function of time. The phase of the slave frequency φ_sappears as sawtooth 802 at the top of the figure while the phase of the master frequency φ_mappears as horizontal line 804. The term “phase” here is not to be interpreted in an analog sense, rather it is applied to the digital contents of the two time of day registers. There registers are incremented by their respective system clocks, so as these clock frequencies slowly drift apart, the register values represent the accumulation of the phase slippage between the two clocks. The phase of the slave frequency is initially declining and will eventually cross the phase of the master frequency. Once this crossing is detected, the clock frequency of the slave will be adjusted.

[0049] The master frequency f_mappears as horizontal line 806 at the bottom of the figure. Slave frequency 808 is shown as a dashed line at the bottom of the figure; it is shown initially 200 PPM greater than the master frequency 806. Vertical dashed lines 810, 812, 814, 816, 818, 820, 822, and 824 indicate the times that the phases of the master and slave signals are compared and, when required, corrections are applied.

[0050] At times 810, 812, 814, and 816 the phase of the slave 802 is greater than the phase of the master 804. During these same intervals, the frequency of the slave 808 is 200 PPM greater than the frequency of the master 806. At synchronization time 818, the phase of the slave 802 is less than the phase of the master 804. When this is detected, the variable dividers in the multistage frequency synthesizer associated with the slave are adjusted to produce a frequency of the slave 808 200 PPM less than the frequency of the master 806. This causes the phase of the slave 802 to rise quickly until at time 820 it is again greater than the master. This causes the frequency of the slave 808 to switch between being 200 PPM less than the master 806 to being 200 PPM greater than the master. During time intervals 822 and 824 the phase of the slave 802 is still greater than the phase of the master 804, so the frequency of the slave 808 remains at 200 PPM greater than the frequency of the master 806.

[0051] As one of ordinary skill in the art will appreciate, once the phase of the slave 802 becomes less than the phase of the master 804, the frequency of the slave 808 will drop to 200 PPM less than the frequency of the master 806 to bring the system back into balance. This continual detection of phase differences and resulting frequency adjustments will keep the time of day registers synchronized during the operation of the computer system. If these adjustments were not made, over a longer period of time the register discrepancies would become so large as to cause system malfunctions as a result of timestamp problems. However, since these synchronization times occur every thousand clock cycles or so in a typical embodiment, the time of day values never shift enough to cause any serious problems. Solutions to this problem in the prior art involved expensive hardware, such as using an external atomic clock, to provide synchronization. Using this invention, two or more time of day registers at different nodes in a multiple processor system can be synchronized with a minimum of additional hardware.

[0052] The description of the present invention has been presented for purposes of illustration and description but is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiment was chosen and described in order to best explain the principles of the invention and the practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.