US 20040080995 A1 Abstract A non-volatile flash memory system counts the occurrences of an event, such as the number of times that individual blocks have been erased and rewritten, by updating a compressed count only once for the occurrence of a large number of such events. A random or pseudo-random number generator outputs a new number in response to individual occurrences of the event, and updates the compressed count when an output of the random number generator matches a predetermined number. The probability of the predetermined number being generated by the random number generator in response to a single event may be varied as the function of some other factor, such as the value of the compressed count, when that provides more useful tracking of the number of events. These techniques also have application to monitoring other types of recurring events in flash memory systems or in other types of electronic systems.
Claims(9) 1. A method of maintaining a compressed count of a number of occurrences of an event that recurs during operation of an electronic system, comprising:
determining whether another event having a random or pseudo-random probability P of occurring in response to individual occurrences of said system event has occurred, and updating a compressed count of the number of occurrences of said system event on those occasions when the randomly or pseudo-randomly occurring event has occurred. 2. A method of maintaining a compressed count of a number of occurrences of an event that recurs during operation of an electronic system, comprising:
generating a random number upon individual occurrences of the event, determining when a generated random number matches at least one predetermined value, and in response to the generated random number matching said at least one predetermined value, updating a compressed count of the number of occurrences of the event within the electronic system. 3. The method of 4. The method of 5. The method of 6. The method of any one of claims 2-5, wherein the method is carried out in an electronic system including non-volatile flash memory and the recurring event includes erasure of an addressed portion of the flash memory. 7. The method of 8. A flash EEPROM system, comprising:
a plurality of blocks of non-volatile memory cells wherein the cells within individual ones of the blocks are simultaneously erasable, a controller including a micro-processor that controls programming of data into addressed blocks of memory cells, reading data from addressed blocks of memory cells and erasing data from one or more of addressed blocks of memory cells at a time, storage provided within the plurality of blocks of memory cells that maintains counts associated with individual ones of the memory cell blocks, a number generator that randomly generates a number in response to one or more of the addressed blocks being erased, and a comparator that causes at least one of the counts associated with one or more addressed blocks being erased to be updated when the generated random number matches a predetermined at least one of possible numbers generated by the random number generator. 9. The system of Description [0001] This invention relates generally to event counting techniques, and, more specifically, to the application of such techniques to semiconductor memory systems, particularly to non-volatile flash electrically-erasable and programmable read-only memories (EEPROMs). [0002] Flash EEPROM systems are being used in a wide variety of applications, particularly when packaged in an enclosed card that is removably connected with a host system. Current commercial memory card formats include that of the Personal Computer Memory Card International Association (PCMCIA), CompactFlash (CF), MultiMediaCard (MMC) and Secure Digital (SD). One supplier of these cards is SanDisk Corporation, assignee of this application. Host systems with which such cards are used include personal computers, notebook computers, hand held computing devices, cameras, audio reproducing devices, and the like. Flash EEPROM systems are also utilized as bulk, mass storage embedded in host systems. [0003] Such non-volatile memory systems include an array of memory cells, peripheral operating circuits and a system controller. The controller manages communication with the host system and operation of the memory cell array to store and retrieve user data. The memory cells are grouped together into blocks of cells, a block of cells being the smallest grouping of cells that are simultaneously erasable. Prior to writing data into one or more blocks of cells, those blocks of cells are erased. User data are typically transferred between the host and memory array in sectors. A sector of user data can be any amount that is convenient to handle, preferably less than or equal to the capacity of the memory block, often being equal to the standard disk drive sector size, which is 512 bytes. [0004] In one commercial architecture, the memory system block is sized to store one sector of user data plus overhead data, the overhead data including information such as an error correction code (ECC) for the user data stored in the block, a count of the number of times that the block has been erased and reprogrammed, defects and other physical information of the memory cell block, and programming and/or erase voltages to be applied to the block. Various implementations of this type of non-volatile memory system are described in the following United States patents and pending applications, each of which is incorporated herein in its entirety by this reference: U.S. Pat. Nos. 5,172,338, 5,602,987, 5,315,541, 5,200,959, 5,270,979, 5,428,621, 5,663,901, 5,532,962, 5,430,859 and 5,712,180, and patent applications Ser. Nos. 08/910,947, filed Aug. 7, 1997, and Ser. No. 09/343,328, filed Jun. 30, 1999. In another commercial architecture, the overhead data for a large number of blocks storing user data are stored together within tables in other blocks. This overhead data includes a count of the number of times that individual user data blocks have been erased and reprogrammed. An example of such a system is described in U.S. patent application Ser. No. 09/505,555, filed Feb. 17, 2000. Yet another type of non-volatile memory system utilizes a larger memory cell block size that stores multiple sectors of user data. [0005] The number of erase/reprogramming cycles experienced by individual memory blocks (their “experience count”) is often maintained within a flash memory system for one or more reasons. One reason is to determine when a block is reaching its end of lifetime, in order to replace it with another block by mapping it out of the system before it fails from overuse. This is described in U.S. Pat. No. 5,043,940, for example, which patent is incorporated herein by this reference. Current commercial floating gate memory cells have a lifetime of from several hundred thousand to one million erase/reprogramming cycles, which is often larger than any of the blocks are cycled in most applications during the useful life of the memory. However, other more reprogramming intensive applications can reach such numbers. Another reason for keeping track of the block experience counts is to be able to alter the mapping of data into the various blocks in order to even out their wear before they reach their ends of lifetime as a way of extending the life of the memory system. Examples of such wear leveling techniques are given in U.S. Pat. No. 6,081,447, which patent is incorporated herein in its entirety by this reference. Yet another reason for maintaining block experience counts is to be able to adjust programming and other operating voltages to take into account changes in characteristics of the memory cells that occur as the number of erase/reprogramming cycles increases. [0006] Rather than keeping track of each occurrence of an event, it is noted only each time a large number of events has occurred. One advantage is that a compressed count R, representative of the number of events A that has occurred, needs to be updated less frequently than if each of the events is counted. Another advantage is that, in a binary counting system, a fewer number of bits are required to maintain a count R that is representative of a larger number of events A. A preferred technique for maintaining the compressed count R includes establishing some probability P that the compressed count R will be updated each time that the event being monitored occurs. This results in updating the compressed count R, on average, once every 1/P number of actual events. This probability is preferably chosen to be as independent of the operation of a system in which the events are occurring as is practical, so that the influence of the system operation upon the frequency of updating the compressed count R is minimized. [0007] In the specific examples described herein, this technique is utilized to monitor the number of some repetitive event that occurs as part of operating an electronic system. A random number generator is preferred for use to determine when the compressed count R of the number of events is updated, a pseudo-random number generator usually being used in practice. A random number is generated when the event occurs, preferably each time the event occurs. One of the random numbers is preferably designated as a trigger to cause the compressed count R to be updated, such as by being incremented to the next number in order. This will occur, on the average, once every N events, where N is the total possible number of distinct random numbers that is generated over time by the random number generator. Rather than each occurrence of the event being counted, therefore, the compressed count R is updated on average once every N events, and that count represents 1/N th the number of events that have occurred, on average. Or, to say it in a different way, the probability P that any one occurrence of an event will result in the compressed count R being updated is 1/N. A product of R and N gives the number of actual events A that has occurred, if that is needed, within a margin of probable error that is proportional to 1/P, which is to say that the likely error goes up as N goes up since P=1/N. [0008] These techniques have particular application to digital memory systems. In the example of non-volatile flash memory systems described in the Background above, updating the compressed count R of an event, such as the erase/reprogramming event, need occur less frequently, so less time is taken away from other operations of the memory. This results in such other operations, such as user data programming, occurring faster. The number of bits required to store the count for each of the blocks is also significantly reduced. Further, the complexity of the operation of the memory system is reduced when each occurrence of the event need not be counted. [0009] In an application of this technique to maintain an experience count (sometimes called a “hot” count) of the number of erasures and reprogramming cycles occurring in a flash memory, the number N is selected to be a small fraction of M, where M is the expected life of the memory in terms of a maximum number of erase/reprogramming cycles that a memory cell can safely experience before there is danger of very inefficient operation or outright failure. Although the resulting compressed count R does not allow knowing exactly the number of events A that have occurred, it's relative accuracy increases as the number of events A grows and particularly when approaching the end M of the memory block's lifetime. The number of bytes of storage space required for the experience count can be significantly reduced since the maximum compressed count R over the life of the memory is M divided by N, rather that being M in the case where every erasure event is counted. Since updating of the compressed count occurs only once for many erasure events, the overall performance of the memory is improved. In addition, the flash memory system is easier to maintain and debug. [0010] The probability P that any particular one of such events will cause the compressed count to be updated need not necessarily be kept the same for the entire time that the events of the memory or other electronic system are being counted but rather can, for example, be varied as a function of the number of events A being monitored. Specifically, if it is desired to maintain a more accurate compressed count R of the erasure/reprogramming events of a flash memory at low values of the actual count A, the probability P is maintained high at the beginning of operation and decreased during the lifetime of the memory as the actual count A becomes large. This is particularly useful, as a specific example, when the compressed count R is being used by the system to control the voltages applied to the memory cells of a particular block during its programming and/or erase, since those voltages are often changed at low levels of the actual experience count A. This ability is provided without having to devote more bits to the storage of the compressed count R for the individual blocks. [0011] Additional aspects, features and advantages of the present invention are included in the following description of specific representative embodiments, which description should be taken in conjunction with the accompanying drawings. [0012]FIG. 1 is a schematic block diagram of a memory system that incorporates event monitoring of the present invention; [0013]FIG. 2 is a block diagram that conceptually illustrates operation of the memory system of FIG. 1 to keep a compressed count of events occurring within it; [0014]FIG. 3 illustrates one form of non-volatile storage within the memory system of FIG. 1 of a compressed count of the events according to the techniques shown in FIG. 2; [0015]FIG. 4 illustrates another form of non-volatile storage within the memory system of FIG. 1 of a compressed count of the events according to the techniques shown in FIG. 2; [0016]FIG. 5 is a flow chart of a specific example of the operation of the memory system of FIG. 1 according to FIG. 2; [0017]FIG. 6 is a table used with the example operation of FIG. 5; [0018]FIG. 7 shows a comparison of two binary numbers that is used in the example of FIG. 5; and [0019]FIG. 8 is another table used in the example of FIG. 5. [0020]FIG. 1 is a diagram of some of the major components of a typical non-volatile memory system. A controller [0021] User data is transferred between the controller [0022] The controller [0023] A logic circuit [0024] The logic circuit [0025] In one typical flash memory implementation, the memory cells of the array are divided into blocks wherein each block is the smallest erasable unit of memory cells, all cells within an individual block being simultaneously erasable. Typically, a number of blocks are erased at the same time, and programming data into the memory array occurs in blocks that have first been erased. In a common example, each block holds 512 bytes of user data plus a number of bytes of overhead data associated with the user data and/or associated with the block of memory cells in which the overhead data are stored. Such a block is formed of two rows of memory cells, in one specific current implementation. In another example, each block holds 32768 (=64×512) bytes of user data plus overhead data. As an alternative to storing the overhead data in the same block as the user data, some or all of the overhead data may be stored in other blocks that are dedicated for that purpose. [0026] One of the items of overhead information associated with each block of current flash memories in a number of flash EEPROM systems is the number of erase/reprogramming cycles the block has experienced. This block experience count is useful for many purposes, the primary ones having been described earlier. When a block experience count is updated to note an event of either the block's erasure or reprogramming, the current experience count stored for the block in the non-volatile memory is first read and stored in a temporary memory, usually volatile memory within the controller. This read count is then updated to represent the occurrence of a subsequent event, such as by incrementing the count by one, and the updated count is then rewritten back into the non-volatile memory block. This involves a significant number of operations that take time and which thus negatively impact upon the performance of the memory system. [0027] According to a principal implementation of the present invention, the experience counts are not updated each time that their associated blocks are erased and reprogrammed. Rather, a compressed count is updated less frequently at an average rate that is related by a proportional constant to the rate of the erase/reprogramming events being counted. For example, if a flash memory has a life of approximately 1,000,000 erase/reprogramming cycles, and the compressed count is updated only once in approximately 4000 cycles, then the updating process occurs only 1/4000 as often over the occurrence of a large number of cycles. The amount of time devoted to updating the compressed experience count during operation of the memory is significantly less than when an experience count is updated by each event. In addition, the number of bits necessary to store the count being maintained is significantly reduced from that required to maintain the actual count, so the space taken in the non-volatile memory to maintain the count is significantly reduced. If an actual count of 1,000,000 cycles is kept, for example, about 3 bytes is required for every block to count each number from 1 to 1,000,000. If an average of only every 4000 [0028] An indication of the actual number of events experienced by a block is always available by multiplying the maintained count by 4000, in this example, but the memory system operating firmware stored in the controller memory [0029] A preferred technique for maintaining the compressed count includes generating a series of random numbers wherein a new number is generated in response to each new event and the compressed count is updated each time the generated random number is equal to a predetermined selected one of those numbers. For example, if a generator of random numbers from 1-4000 is employed, where a new number is generated each time the event occurs, and a particular number, say 2750, is selected to update the maintained count when it occurs, the compressed count will be updated each time the random number generator output is equal to 2750. On average, this will occur once each 4000 events. Although there is not a precise relationship between the number of events that have occurred and the compressed count, there is a close correlation, particularly after a large number of events has occurred. It has been found that the accuracy of the compressed count is sufficient for the purposes of the example being described, particularly after several hundred thousand events have occurred. This is when the experience count information becomes quite useful, in the example application being described, for determining when individual blocks of memory need to be replaced. [0030]FIG. 2 conceptually illustrates this process implemented in the non-volatile memory system of FIG. 1, as an example. A random number generator [0031] One number contained within the set of numbers associated with the random number generator is stored in a predetermined location within the system in a non-volatile manner, as indicated by [0032] In one embodiment, the counts are maintained for a number of user data blocks in a single one of many reserved blocks, such as a reserved block [0033] In another embodiment, instead of the counts being maintained in a separate reserved block [0034] Yet, in another embodiment, the counts are stored in a block that is exclusively devoted to storing the counts and no other kind of overhead data. [0035] The random number generator [0036] The random number generator, to provide more detail, uses a 32 bit shift register formed in the controller RAM [0037] It will be noted that the probability P may be generated by some means other than the random number generator [0038] Upon the memory system being initiated, as the result of being powered up from an un-powered condition, the random number generator [0039] There are many alternative seed sources that can be employed. One is to store, in a non-volatile manner, the last value [0040] Yet another alternative technique for generating a seed involves reading user data in a block of the memory which can be read in a normal manner. But to assure a higher degree of randomness, in case the data does not change between initializations, that reading is performed, in one specific implementation, with margin threshold levels that are separated by amounts far in excess of those normally used for reading data. This is intended to assure, due to the marginality of the readings, that many errors will occur in reading that data, preferably in a manner that at least some of the same data are read differently at different times. To further increase the randomness of the seed, an address of a second block may be generated from this intentionally erroneously read data, and the data in the second block is read in the same manner that is likely to erroneously read the data. This can be continued for a further number of cycles if additional assurances of randomness are desired. [0041] In the description given above, it has been assumed that the probability P of a match occurring remains the same throughout the life of the memory system. There can be applications, however, where it is desired to vary the probability P in some manner, such as in response to some relevant condition that changes or to a related event that occurs. That probability can be altered, if desired, by changing the number of predetermined numbers in the storage [0042] As an alternative to changing the number of stored predetermined numbers that are compared in order to alter the probability, the number of bits of a single number in the storage [0043] One specific application in a flash memory system of varying the probability of a match occurring is described with respect to the flowchart of FIG. 5. This operating method increments the compressed count R more often at lower numbers of R which usually correspond to lower numbers of the actual counts A than at higher numbers, thus resulting in the compressed count R more accurately representing the actual count A at the lower numbers. The resolution of the compressed count R at lower numbers is thus improved. This can be accomplished without having to increase the number of bits required to store the compressed count R by initially using a high value of P at low values of R, and proceeding to lower values of P (which can become lower than 1/4096 and in fact can become as low as 1/32768) at higher values of R. By using this approach, an integer R ranging from 0 to 255 can represent counts as high as 1 million for the high count range, and as low as single digit numbers for the lowest count range. In the specific flash EEPROM system example described herein, the total number of binary random number bits that are compared to the same number of bits of the pre-designated number can be designed to be a function of the value of the compressed count R of each block being erased. [0044] In the specific example being described with respect to FIGS. [0045] Referring to FIG. 5, a first step [0046] The number of bits of the numbers [0047] The number of bits of the numbers [0048] Although the examples being described include incrementing various numbers, one or more could be decremented instead, from a high value to a low value. It is not usually important how the compressed count R is updated to record the fact of a positive comparison between the numbers [0049] A table such as that illustrated in FIG. 8 is also optionally included as part of a software package used by failure analysis engineers in order to relate the compressed count R to the actual count A in those cases where an estimate of A needs to be known. (In most cases, the memory system firmware can operate from the count R itself since its relationship to the count A is a known one.) For a first group [0050] The relationship between the counts R and A are usually not those roughly determined as above, particularly as the value of R becomes high, because of the incremental contribution of a combination of the effects of different probability values in the different ranges of R represented by prior groups [0051] The error numbers in the right-hand column of FIG. 8 are also best determined by such an empirical technique, when it is desired to include them. The counts R and A are maintained during a large number of cycles of a number of memory blocks, and the differences among the different blocks are statistically expressed in some manner, such as by a standard deviation of a population of such differences. This column can be useful for the purposes of evaluating various tradeoffs between desired precision and necessary memory space required to store the counts, and diagnoses by the user of the memory system, but will usually not be used by the controller [0052] Although the examples described herein are for maintaining a count of the number of erase/rewrite cycles experienced by blocks of a flash EEPROM system, these techniques can also be applied with similar advantages to counting other events in such a system, particularly when a separate count of an event is maintained for each block or group of blocks of memory cells. One example is to count the number of times that individual blocks are subjected to margin scanning data recovery techniques, as an indication of some problem with those blocks. Another example is to count the number of times that an error correction code (ECC) was actually engaged in order to recover the user data of an otherwise unreadable sector. Another is to keep track of the number of times that blocks need to have the data therein refreshed to their proper margin levels by scrubbing in order to compensate for disturbances that have occurred over time. In these and other situations, a large number of such events experienced by a block, a sub-block (sector), or a group of blocks gives an indication that there is some problem with them. This information can be used by the memory controller to replace the block, sub-block, or the group of troubled blocks, or take some other remedial action. [0053] Further, the counting techniques described above are not limited to use with flash EEPROM or other non-volatile memory systems. These techniques have application in any electronic system where it is necessary or desirable to keep a count of one or more events occurring in the course of the operation or by the use of the system. [0054] Although the various aspects of the present invention have been described with respect to specific exemplary embodiments, it will be understood that the invention is entitled to protection within the full scope of the appended claims. [0055] A New Compressed Stochastic Integer Event Counter Scheme [0056] This idea is applicable to any situation when a record of the number of times a certain event has occurred must be kept. The usual way of keeping track of the number of times an event has occurred is to devise a counter that is incremented every time an event takes place. If the maximum number of times the event can possibly occur is M times, then to store this information the counter requires N=log [0057] The Simple Approach Using a Fixed Probability of Incrementation [0058] The new idea allows the compression of the 3 hot count Bytes per sector to 1 hot count Byte per sector. If one is willing to give up precision, then in its simplest form the new hot count scheme is as follows: using a 12 bit pseudo-random number generator, one can generate a probability p (in this case p=1/4096=0.0002441). One way of generating this probability is by generating a random 12 bit binary integer on every occasion that a sector is being programmed, or erased. If this 12 bit integer matches a particular 12 bit integer (say 101111010001), then, and only then, will the counter corresponding to the sector that is about to be written be incremented. Note that 2 [0059] Now, we have to distinguish between the actual count, A, which is the actual number of times a given sector has been programmed, and the representation of this count, R, which is roughly 4096 times smaller than A. What we record and keep track of is R=r. The value of R at any given time will not exactly determine the value of A=a, but will give a ball park estimate of the value of A. When R is small, then R is a very poor indicator of A, but as R becomes larger, then R becomes a more accurate indicator of A in a relative sense. Here “R”, and “A” represent random variables, whereas “r”, and “a” represent certain integer values that these random variables can take. [0060] We can define E as the event of R being incremented. Then: [0061] Probability (E)=Pr(E)=p, and Probability (not E)=Pr({overscore (E)})=1−p [0062] “r”=number of times that E occurs in “a” trials=random variable with probability density function f(r|a) given by the binomial distribution: [0063] {The mean value of R}=Exp(R)=a·p, and [0064] {The standard deviation of R}=SD(R)={square root}[a·p·(1−p)] [0065] The above formulas are well known for the binomial distribution. [0066] For the purposes of this hot count scheme we need to know the expectation value of A given some known value of R, and the standard deviation of A given some known value of R. This is because at any given time only the value of R is stored and can be retrieved, and no knowledge of the exact value of A has been retained. [0067] One is tempted to write Exp(A|R=r)=r/p. While this may be correct, to be sure, Bayes Theorem has to be invoked to derive an expression for Pr(R=r|A=a)=f(a|r), and then this probability density function can be used to calculate Exp(A|R=r), and SD(A|R). Bayes Theorem states the following:
[0068] The denominator of the right hand side of the above equation can be rewritten by using the following identity:
[0069] where we know the actual count A can never be smaller than the representation R (hence the summation starts from r), and we assume that the actual count A can never exceed some large number M (say: M=1000000 or M=10000000). Therefore {A=r, A=r+1, A=r+2, . . . , A=M} forms a partition. Also, by the definition of conditional probability:
[0070] There is a theorem in statistics called “The Admissibility of Uniform Distribution for Bayesian Estimates” that states in cases such as we have here, it is safe to assume that the random variable A is uniformly distributed. In other words, if we assume that the actual count can never exceed M, and we do not know the value of the representation R, then the value of A at some random instance in time can be any integer from 0 to M, with each integer value having the same likelihood as any other. Therefore, Pr(A=s)=constant=(1/(M+1)), and we can pull this constant out of the summation:
[0071] and because A is uniformly distributed we know that Pr(A=s)=Pr(A=a) for any value of s, and any value of a:
[0072] where:
[0073] Therefore:
[0074] The average value of A given R=r is given by:
[0075] The variance of (A|R=r) is:
[0076] The Standard Deviation of A given R=r is:
[0077] As is apparent from the above formulas, even in this simple case where p=constant, the calculations require computer programs. However, Wald's Equation (See the book STOCHASTIC PROCESSES by Sheldon M. Ross, from John Wiley & Sons (1983) pp 59) can be used to readily arrive at the following expression for: Exp(A|R=r)=Exp(R|A=a)/p=r/p. Therefore, in the case that p=1/4096, then: Exp(A|R=r)=4095·r [0078] Another method of calculating both the expectation values, and the standard deviations of A, given some value of R=r is by performing Monte Carlo Simulations. These simulations are particularly useful when a more complicated algorithm is adopted such as the “r dependent p value algorithm” discussed in the next section. Also since any random number generating scheme really generates a pseudo-random number with a distribution that may not be perfectly uniform, the best platform on which to perform the Monte Carlo simulation consists of the same processor and firmware that will go into production. In this way, the imperfections of the random number generating scheme are also incorporated into the simulation. [0079] The Monte Carlo simulation for this simple case will consist of 2 phases. The first phase consists of two loops, with one loop nested inside the other loop. The inner loop will increment the actual count “a” by one upon every pass. Also every time the integer “a” is incremented, a 12 bit random binary integer, b, is generated, if and only if this integer becomes equal to “101111010001”, then the representation “r” is incremented. As discussed previously the probability for such a match is 1 in 4096. Note that both “a” and “r” are set to zero within the outer loop, and outside the inner loop. This allows every trial to start with a=0 and r=0. The inner loop is repeated as long as r<256. On the average the inner loop will be run 1 million times. The inner loop simulates what happens to the hot count of a single sector as the sector is cycled for roughly one million times (until r saturates at 255 [0080] A More Complex Approach Using a Variable Probability Value of Incrementation [0081] A more complex algorithm can be adopted in order to keep the ratio of (the standard deviation of A given R=r) to (the expectation value of A given R=r) more or less constant for different values of “r”. This is useful because it limits the relative uncertainty in the value of “a” based on the knowledge of the value of “r”. In this new scheme, the probability “p” that “r” gets incremented is related to the current value of “r”. [0082] If 0≦r<16, then p=1/2 [0083] If 16≦r<32, then p=1/2 [0084] If 32≦r<48, then p=1/2 [0085] If 48≦r<64, then p=b [0086] . . . [0087] . . . [0088] . . . [0089] If 240≦r<256, then p=1/2 [0090] “r” can be represented as a simple count from 00000000 to 11111111 in binary form, with a decimal equivalence being from 0 to 255. In Hex, the value of “r” can range from 00h to FFh. Each user of the hot count “r” will be provided with a table that provides a mean value of A, and a standard deviation of A for each of the 256 different values of “r”. Alternatively, “r” can be represented in the form of a 4 bit mantissa and a 4 bit exponent, as discussed later. But this later representation is cumbersome at best. [0091] The Monte Carlo simulation for this scheme is very similar to the previously discussed simple scheme, with the difference that now the value of “p” will depend on the current value of“r”, as explained above. [0092] A New Hot Count and Counters Scheme for a Specific Memory [0093] Each page (64 sectors) will have a single, Byte long, hot count that will be stored out of the page itself, and in a table in another page. The hot count will be updated in the controller RAM every time it is decided that the count requires incrementing. Every time a page is erased a decision will be made to either increment the corresponding hot count or not to increment it. In this scheme the chances of incrementing the hot count will roughly be inversely proportional to the present value of the count. [0094] The table containing the hot counts will have to be about 5000 Bytes in order to contain the hot counts of about 5000 pages. Assuming 2 [0095] Every time a read or write operation crosses over a partition boundary, then one hot count table sector is updated, and all hot count table sectors are read and processed so that wear leveling, scrubbing, and retirement activity may now be performed. The logical order of the hot count Bytes represents the physical order of the corresponding pages. Also, The logical order of the hot count tables represents the physical order of the corresponding partitions. [0096] Updating of a hot count sector will normally not even require an erase, because as with any other sector, the updating of a sector only requires writing the sector data in a new pre-erased sector, and updating the map to reflect where the most recent version of the data resides. Since a partition consists of 64 pages, the hot counts corresponding to the partition will be 64 bites only. In this scheme the single Byte hot count of each page is incremented only 256 times during a million cycles. So, even if we would update the hot count of each page as soon as it was incremented, and the sector containing the hot counts of 512 pages were never moved around, then the hot count table sector should be updated 512*256=131072 times. This assumes that each and ever) one of these 512 pages have been cycled 1 million times. If user sectors can tolerate 1 million cycles, then hot count sectors too can tolerate 131 thousand cycles. [0097] The Hot Count Scheme: [0098] The 8 bit count corresponding to each page will consist of 4 bits of mantissa, M, and 4 bits of exponent, E. So if the hot count reads: 01010010, then M=0101 [0099] Now the actual hot count in base 10 is: N(E) [0100] Alternatively, the hot count in base 2 is: N(E)*(10)**E [0101] For counts from 0 to 15 every erase of the page will increment the hot count with 100% certainty. For counts from 16 to 46 every erase has a 50% chance of incrementing the hot count. [0102] In general the probability, p, of incrementing is a function of E given by: p=1/(2**E) [0103] The following are values of A(E): [0104] A(0000)=0000, [0105] A(0001)=1000, [0106] A(0010)=1100, [0107] A(0011)=1110, [0108] A(0100)=1111, [0109] A(0101)=1111, [0110] A(0110)=1111, [0111] A(0111)=1111, [0112] . . . [0113] A(1111)=1111 [0114] The largest possible number, L, is given by, M=1111, and E=1111: [0115] L=(1111+1111)*((10)**(1111)=1015792 [0116] The smallest p is 1/32768. [0117] This same methodology can be applied to counts of counter activity also. Please see the attached Excel worksheet for the details of this scheme. Because cycled cells age more slowly when they have been cycled many times, as compared to the beginning of life, this proportionate hot count will have enough resolution for both wear leveling and retirement purposes. Referenced by
Classifications
Rotate |