Patent Application
20080222364
1. Field of the Invention
The present invention generally relates to computer systems having multiprocessor architectures and, more particularly, to a novel multi-processor computer system for processing memory accesses requests and the implementation of cache coherence in such multiprocessor systems.
2. Description of the Prior Art
To achieve high performance computing, multiple individual processors have been interconnected to form multiprocessor computer system capable of parallel processing. Multiple processors can be placed on a single chip, or several chips—each containing one or several processors—interconnected into a multiprocessor computer system.
Processors in a multiprocessor computer system use private cache memories because of their short access time (a cache is local to a processor and provides fast access to data) and to reduce number of memory requests to the main memory. However, managing caches in multiprocessor system is complex. Multiple private caches introduce the multi-cache coherency problem (or stale data problem) due to multiple copies of main memory data that can concurrently exist in the multiprocessor system.
Small scale shared memory multiprocessing system have processors (or groups thereof) interconnected by a single bus. However, with the increasing speed of processors, the feasible number of processors which can share the bus effectively decreases.
The protocols that maintain the coherence between multiple processors are called cache coherence protocols. Cache coherence protocols track any sharing of data block between the processors. Depending upon how data sharing is tracked, cache coherence protocols can be grouped into two classes: 1) Directory based and 2) Snooping.
In directory based approach, the sharing status of a block of physical memory is kept in just one location called the coherency directory. Coherency directories are generally large blocks of memory which keep track of which processor in the multiprocessor computer system owns which lines of memory. Disadvantageously, coherency directories are typically large and slow. They can severely degrade overall system performance since they introduce additional latency for every memory access request by requiring that each access to the memory go through the common directory.
In a snooping approach, no centralized state is kept, but rather each cache keeps the sharing status of data block locally. The caches are usually on a shared memory bus, and all cache controllers snoop (monitor) the bus to determine whether they have a copy of the data block requested. A commonly used snooping method is the “write-invalidate” protocol. In this protocol, a processor ensures that it has exclusive access to data before it writes that data. On each write, all other copies of the data in all other caches are invalidated. If two or more processors attempt to write the same data simultaneously, only one of them wins the race, causing the other processors' copies to be invalidated.
To perform a write in a write-invalidate protocol based system, a processor acquires the shared bus, and broadcasts the address to be invalidated on the bus. All processors snoop on the bus, and check to see if the data is in their cache. If so, these data are invalidated. Thus, use of the shared bus enforces write serialization.
Disadvantageously, every bus transaction in the snooping approach has to check the cache address tags, which could interfere with CPU cache accesses. In most recent architectures, this is typically reduced by duplicating the address tags, so that the CPU and the snooping requests may proceed in parallel. An alternative approach is to employ a multilevel cache with inclusion, so that every entry in the primary cache is duplicated in the lower level cache. Then, snoop activity is performed at the secondary level cache and does not interfere with the CPU activity.
With the increasing number of processors on a bus, snooping activity increases as well. Unnecessary snoop requests to a cache can degrade processor performance, and each snoop requests accessing the cache directory consumes power. In addition, duplicating the cache directory for every processor to support snooping activity significantly increases the size of the chip. This is especially important for systems on a single chip with a limited power budget.
What now follows is a description of prior art references that address the various problems of conventional snooping approaches found in multiprocessor systems.
Particularly, U.S. Patent Application US2003/0135696A1 and U.S. Pat. No. 6,704,845B2 both describe replacement policy methods for replacing entries in the snoop filter for a coherence directory based approach including a snoop filter. The snoop filter contains information on cached memory blocks—where the cache line is cached and its status. The U.S. Patent Application US2004/0003184A1 describes a snoop filter containing sub-snoop filters for recording even and odd address lines which record local cache lines accessed by remote nodes (sub-filters use same filtering approach). Each of these disclosures do not teach or suggest a system and method for locally reducing the number of snoop requests presented to each cache in a multiprocessor system. Nor do they teach or suggest coupling several snoop filters with various filtering methods, nor do they teach or suggest providing point-to-point interconnection of snooping information to caches.
U.S. Patent Applications US2003/0070016A1 and US2003/0065843A1 describe a multi-processor system with a central coherency directory containing a snoop filter. The snoop filter described in these applications reduces the number of cycles to process a snoop request, however, does not reduce the number of snoop requests presented to a cache.
U.S. Pat. No. 5,966,729 describes a multi-processor system sharing a bus using a snooping approach for cache coherence and a snoop filter associated locally to each processor group. To reduce snooping activity, a list of remote processor groups “interested” and “not-interested” in particular cache line is kept. Snoop requests are forwarded only to the processor groups marked as “interested” thus reducing the number of broadcasted snoop requests. It does not describe how to reduce the number of snoop requests to a local processor, but rather how to reduce the number of snoop requests sent to other processor groups marked as “not interested”. This solution requires keeping a list with information on interested groups for each line in the cache for a processor group, which is comparable in size to duplicating the cache directories of each processor in the processor group thus significantly increasing the size of chip.
U.S. Pat. No. 6,389,517B1 describes a method for snooping cache coherence to allow for concurrent access on the cache from both the processor and the snoop accesses having two access queues. The embodiment disclosed is directed to a shared bus configuration. It does not describe a method for reducing the number of snoop requests presented to the cache.
U.S. Pat. No. 5,572,701 describes a bus-based snoop method for reducing the interference of a low speed bus to a high speed bus and processor. The snoop bus control unit buffers addresses and data from the low speed bus until the processor releases the high speed bus. Then it transfers data and invalidates the corresponding lines in the cache. This disclosure does not describe a multiprocessor system where all components communicate via a high-speed bus.
A. Moshovos, G. Memik, B. Falsafi and A. Choudhary, in a reference entitled “JETTY: filtering snoops for reduced energy consumption in SMP servers” (“Jetty”) describe several proposals for reducing snoop requests using hardware filter. It describes the multiprocessor system where snoop requests are distributed via a shared system bus. To reduce the number of snoop requests presented to a processor, one or several various snoop filters are used.
However, the system described in Jetty has significant limitations as to performance, supported system and more specifically interconnect architectures, and lack of support for multiporting. More specifically, the approach described in Jetty is based on a shared system bus which established a common event ordering across the system. While such global time ordering is desirable to simplify the filter architecture, it limited the possible system configurations to those with a single shared bus. Alas, shared bus systems are known to be limited in scalability due to contention to the single global resource. In addition, global buses tend to be slow, due to the high load of multiple components attached to them, and inefficient to place in chip multiprocessors.
Thus, in a highly optimized high-bandwidth system, it is desirable to provide alternate system architectures, such as star, or point-to-point implementations. These are advantageous, as they only have a single sender and transmitter, reducing the load, allowing the use of high speed protocols, and simplifying floor planning in chip multiprocessors. Using point to point protocols also allows to have several transmissions in-progress simultaneously, thereby increasing the data transfer parallelism and overall data throughput.
Other limitations of Jetty include the inability to perform snoop filtering on several requests simultaneously, as in Jetty, simultaneous snoop requests from several processors have to be serialized by the system bus. Allowing the processing of several snoop requests concurrently would provide a significant increase in the number of requests which can be handled at any one time, and thus increase overall system performance.
Having set forth the limitations of the prior art, it is clear that what is required is a system incorporating snoop filters to increase overall performance and power efficiency without limiting the system design options, and more specifically, methods and apparatus to support snoop filtering in systems not requiring a common bus.
Furthermore, there is a need for a snoop filter architecture supporting systems using point-to-point connections to allow the implementation of high performance systems using snoop filtering.
There is a further need for the simultaneous operation of multiple snoop filter units to concurrently filter requests from multiple memory writers to increase system performance.
There is further a need to provide novel, high performance snoop filters which can be implemented in a pipelined fashion to enable high system clock speeds in systems utilizing such snoop filters.
There is an additional need for snoop filters with high filtering efficiency transcending the limitations of prior art.
It is therefore an object of the present invention to provide a simple system and method for reducing the number of snoop requests presented to processors in cache coherent multiprocessor systems.
It is a further object of the present invention to provide a system and method for effectively reducing the local snooping activity by the addition of a simple hardware device locally to each processor which filters out incoming snoop requests. The snoop filter described is significantly smaller than a mirrored cache directory, and identifies most of the snoop requests which would result in misses, while never filtering out a snoop requests for lines that are locally cached. Reducing the number of snoop requests per processor increases system performance and reduces power.
Thus, according to a first aspect of the invention, there is provided a snoop filtering system and method in a computing environment having multiple processing units, each processing unit having one or more cache memories associated therewith, the system comprising:
According to this embodiment, each of the one or more parallel operating sub-filter elements generates a signal indicating whether a snoop request is to be forwarded to the associated processor or not forwarded, and the snoop filter means additionally comprises a means responsive to each signal generated from said sub-filter element for deciding whether a snoop request is to be forwarded or discarded.
According to a second aspect of the invention, there is provided a snoop filtering system for supporting cache coherency in a computing environment having multiple processing units, each processing unit having associated one or more cache memory system operatively coupled therewith, the system comprising: a plurality of interconnected snoop filter units, each snoop filter unit corresponding to and in communication with a respective processing unit, with each snoop filter unit comprising;
whereby a number of snoop requests forwarded to a processing unit is reduced thereby increasing performance of the computing environment.
According to this aspect of the invention, the plurality of interconnected snoop filter units are connected by point-to-point interconnect means comprising communication links for directly connecting a memory writing source to a corresponding receiving means of each snoop filter unit associated with each processing unit of said multiprocessor computing environment.
Further, in one example implementation, each cache memory system comprises a first level cache and a second level cache operatively interconnected with said processor unit and a shared main memory storage device. The snoop filter unit is operatively coupled to a first level cache for filtering out snoop requests at said first cache level. Alternately, the snoop filter unit is operatively coupled to a second level cache for filtering out snoop requests at the second cache level.
In one example implementation, the cache memory system is organized as a write-through cache, wherein snooping only occurs on write requests. Accordingly, the processing unit implements a cache coherence protocol for invalidating a local data copy as a result of a snoop action.
In another example implementation, the cache memory system is organized as a write-back cache, wherein snooping only occurs on both read and write requests. Thus, the processing unit implements a cache coherence protocol for determining if remote caches have the most recent data copy in relation to a main memory.
The objects, features and advantages of the present invention will become apparent to one skilled in the art, in view of the following detailed description taken in combination with the attached drawings, in which:
a) and 8(b) depict high level schematics of two alternative embodiments of the snoop block according to the present invention;
a) and 19(b) depict exemplary cache wrap detection logic circuitry (registers and comparator) for an N-way set-associative cache;
Referring now to drawings, and more particularly to
To reduce the number of snoop requests presented to a processor, and thus to reduce the impact of snooping on processor and system performance, and to reduce power consumed by unnecessary snoop requests, a snoop filter 140a, . . . , 140n is provided for each respective processor core 100a, . . . , 100n in the multiprocessor system 10. For transferring snooping requests, the preferred embodiment does not use the system bus 150, as typically found in prior art systems, but rather implements a point-to-point interconnection 160 whereby each processor's associated snoop filter is directly connected with each snoop filter associated with every other processor in the system. Thus, snoop requests are decoupled from all other memory requests transferred via the system local bus, reducing the congestion of the bus which is often a system bottleneck. All snoop requests to a single processor are forwarded to the snoop filter 140a, . . . , 140n, which comprises several sub-filters with the same filtering method, or with several different filtering methods, or any combination of the two, as will be described in greater detail herein. The snoop filter processes each snoop request, and presents only a fraction of all requests which are possibly in the processor's cache to the processor.
For each processor, snoop requests are connected directly to all other processors' snoop filters using a point-to-point interconnection 160. Thus, several snoop requests (resulting from write and invalidate attempts) from different processors can occur simultaneously. These requests are no longer serialized, as in the typical snooping approach using the system bus, where this serialization is performed by the bus. That is, multiple snoop requests can be processed in the snoop filter concurrently, as will be described herein in further detail. As a processor has only one snoop port, the snoop requests not filtered out by a snoop filter will be serialized in a queue to be presented to the processor. However, the number of requests passed to the processor is much less than the pre-filtered number of all snoop requests, reducing the impact of cache coherence implementation on system performance.
To prevent queue overflowing condition of the queues contained in the snoop filter block, a token-based flow control system is implemented for each point to point link to limit the number of simultaneously outstanding requests. According to the token-based flow control, each memory writer can send the next write request—which also initiates snoop requests to all other processor units and accompanied snoop filter blocks—only if it has tokens available for all ports of the snoop filter blocks it has a direct point-to-point connection. If there are no tokens available from at least one of the remote ports it is connected to, no snoop requests can be sent out from this memory writer until at least one token from the said snoop filter port gets available again.
Referring now to
In operation, the port snoop filters 400a, . . . , 400n process the incoming snoop requests and forward a subset of all snoop requests to a respective snoop queue 420a, . . . , 420n having one queue associated with each snoop port. A queue arbitration block 422 is provided that arbitrates between all the snoop queues 420 and serializes all snoop requests from the snoop queues 420 fairly. Logic is provided to detect a snoop queue overflow condition, and the status of each queue is an input to a snoop token control unit 426 that controls flow of snoop requests from the remote memory writers. A memory writer—being a processor or a DMA engine—can submit a write to the memory and a snoop request to all snoop filters only if it has a token available from all snoop filters. The only snoop filter from which a processor does not need a token available to submit a write is its own local snoop filter. This mechanism ensures that the snoop queues do not overflow. From the snoop queue selected by arbiter 422, snoop requests are forwarded to the processor via a processor snoop interface 408.
According to the preferred embodiment, the snoop cache block 440 filters the snoop requests 410 using an algorithm which is based on the temporal locality property of snoop requests, meaning that if a single snoop request for a particular location was made, it is probable that another request to the same location will be made soon. The snoop cache monitors every load made to the local cache, and updates its status, if needed. The stream register check block 444 filters snoop requests 410 using an algorithm that determines a superset of the current local cache content. The approximation of cache content is included in the stream registers block 430 (
In the preferred embodiment, only the three filtering units implementing the algorithms above described are included in a port snoop filter, but one skilled in the art will appreciate that any other number of snoop filter units can be included in a single port snoop filter, or that some other snoop filter algorithm may be implemented in the port snoop filter, or a combination of snoop algorithms can be implemented, without departing from the scope of the invention.
a) and 8(b) depict high level schematics of two alternative embodiments of the snoop filter block 310 of
The port snoop filters 460a, . . . , 460n filter the incoming snoops and forward the appropriate unfiltered subset of snoop requests into the respective queues 466a, . . . , 466n and the queue arbitration block 468. Here, the snoop requests are serialized and presented to a next snoop filter 470, which handles inputs from all remote memory writers. This shared snoop filter 470 processes all snoop request presented and forwards a subset of all requests to the snoop queue 472. From the snoop queue 472, snoop requests are forwarded to the processor via the processor snoop interface 474. It should be understood that it is possible to have multiple or no shared snoop filters 470 instead of the configuration shown in
Similarly,
Referring now to
All snoop cache devices 502a, . . . , 502n also receive read addresses and requests 512 from the local processor, and compare the memory read access addresses to the entries in the snoop cache 502a, . . . , 502n. If a request matches one of the entries in the snoop cache, this entry is removed from the snoop cache, as now the cache line is going to be located in the processor's first level cache. In the preferred embodiment, multiple snoop caches operating in parallel are used, each keeping track of snoop requests from a single remote memory writer. After filtering, a fraction of unfiltered snoop requests can be forwarded to the next port snoop filter, or they can be queued for one or more shared snoop filters, or they are placed in the snoop queue of the processor interface, depending on the embodiment.
It is understood that a single snoop cache device 502 includes an internal organization of M cache lines (entries), each entry having two fields: an address tag field, and a valid line vector. The address tag field of the snoop cache is typically not the same as the address tag of the L1 cache for the local processor, but it is shorter by the number of bits represented in the valid line vector. Particularly, the valid line vector encodes a group of several consecutive cache lines, all sharing the same upper bits represented by the corresponding address tag field. Thus, the n least significant bits from an address are used for encoding 2n consecutive L1 cache lines. In the extreme case when n is zero, the whole entry in the snoop cache represents only one L1 cache line. In this case, the valid line vector has only one bit corresponding to a “valid” bit.
The size of the address tag field in the snoop cache is determined by the size of the L1 cache line and the number of bits used for encoding the valid line vector. In an example embodiment, for an address length of 32 bits (31:0), an L1 cache line being 32 bytes long, and a valid line vector of 32 bits, address bits (31:10) are used as the address tag field, (bit 31 being the most significant), address bits (9:5) are encoded in the valid line vector, and address bits (4:0) are ignored because they encode the cache line byte offset. As an illustration, three snoop caches for three different memory writers (N=3) are listed below, each snoop cache having M=4 entries, with address tag field to the left, and with 5 bits from the address used to encode the valid line vector to track 32 consecutive cache lines:
Snoop Requests Source 1
Entry 1: 01c019e 00000000000000000001000000000000
Entry 2: 01c01a0 00000000000000000000000100000000
Entry 3: 01c01a2 00000000000000000000000000010000
Snoop Requests Source 2
Entry 1: 01c01e3 00010000000000000000000000000000
Entry 2: 01c01e5 00000001000000000000000000000000
Entry 3: 01c01e7 00000000000100000000000000000000
Snoop Requests Source 3
Entry 1: 01c0227 00000000000000000001000000000000
Entry 2: 01c0229 00000000000000000000000100000000
Entry 3: 01c022b 00000000000000000000000000010000
In this example, entry 1 of the source 1 snoop cache has recorded that address 01c019ec hexadecimal has been invalidated recently and cannot possibly be in the L1 cache. Therefore, the next snoop request to the same cache line will be filtered out (discarded). Similarly, entry 4 of the source 1 snoop cache will cause snoop requests for cache line addresses 01407ff7 and 01407ff8 to be filtered out.
Referring now to
At step 606, a determination is made as to whether both the “tag” field matches and the corresponding bit in the valid line vector is set. If both the “tag” field matches and the corresponding bit in the valid line vector is set, the snoop request is guaranteed not to be in the cache as indicated at step 606. Thus, this snoop request is not forwarded to the cache; it is filtered out as indicated at step 608.
Otherwise, if the address “tag” field hits in the snoop cache but the bit in the valid line vector is not set or, alternately, if the tag does not hit in the snoop cache, this indicates that the line may be in the cache. Consequently, the snoop request is forwarded to the cache by placing it into a snoop queue as indicated at step 612. This snoop request is also added as a new entry to the snoop cache as shown at step 610.
Referring now to
In yet another embodiment, the new information is not added into the snoop cache based on the hit or miss of a snoop request in the snoop cache only, but instead, the addition of new values—being whole snoop cache lines or only setting a single bit in a valid line vector—is based on the decision of the decision logic block 450 (
Referring now to
Returning to step 632, if it is determined that the comparison of the address tag with all snoop caches results in one or more hits, the information has to be removed from all snoop caches for which it was hit. Thus, at step 634, the appropriate low order bits of the memory read address are decoded into a valid line vector, and is matched against the valid line vector of the snoop cache entry that was hit as indicated in step 635. Proceeding now to step 636, it is determined whether the unique bit set in the read address vector is also set in the valid line vector of the snoop cache. If there is no such valid line vector hit (regardless of the address tag field hit), this memory address is not filtered out and nothing has to be changed in the particular snoop cache. Thus, the control flow proceeds to step 640 to check if all address tag hits have been processed, and if not, the process returns to step 635.
If, however, it is determined at step 636 that the read address vector hits in the valid line vector, then the read address is being filtered out. The corresponding valid line vector bit has to be cleared since the memory read address is going to be loaded into the first level cache. This clearing of the corresponding bit in the valid line vector is performed at step 638. If after removing the corresponding bit from the valid line vector the number of bits set of the valid line vector becomes zero, the address tag field is further removed from the snoop cache causing the entry to be empty. As next indicated at step 640, the same process of checking for the valid line vector bit, its clearing, and clearing of the address tag—if necessary—is repeated for all snoop caches which hit the memory read request which was miss in the local L1 cache. This condition that all hit address tag lines have been processed is checked at step 640. Once all of the cache lines have been checked, the process returns to step 630.
In yet another embodiment, the local memory request is compared to all address tags in all snoop caches simultaneously. Concurrently, the valid line vector encoding of the local memory request may be compared with all valid line vectors in all snoop caches in which there were hits simultaneously. Then, these two results—address tag hit and valid line vector hit—can be combined to determine all snoop cache lines from which the corresponding valid line vector bit has to be removed, and all these bits can be removed from the hitting cache lines from all snoop caches simultaneously.
Referring now to
The heart of the stream register filter is the stream registers 700 themselves. One of these registers is updated every time the cache loads a new line, which is presented to the stream registers with appropriate control signals 716. Logic block 704 in
In operation, snoop requests received from one of the N remote processors arrive as signals 714 shown in the right-hand side of
The arbitrate and multiplex logic block 710 simply shares the snoop interface of the cache between the N snoop request queues 703 in a fair manner, guaranteeing forward progress for all requests.
A description of how a single stream register is updated is now provided. A stream register actually comprises a pair of registers, the “base” and the “mask”, and a valid bit. The base register keeps track of address bits that are common to all of the cache lines represented by the stream register, while the corresponding mask register keeps track of which bits these are. The valid bit simply indicates that the stream register is in use and should be consulted by the snoop check logic 702 when deciding whether to filter a remote snoop request 714. In order to understand the examples in the following description, consider an address space of 232 bytes with a cache line size of 32 bytes. In this case, a cache line load address is 27 bits in length, and the base and mask registers of the stream registers are also 27 bits in length.
Initially, the valid bit is set to zero, indicating that the stream register is not in use, and the contents of the base and mask register is irrelevant. When the first cache line load address is added to this stream register, the valid bit is set to one, the base register is set to the line address, and all the bits of the mask register are set to one, indicating that all of the bits in the base register are significant. That is, an address that matches the address stored in the base register exactly is considered to be in the cache, while an address differing in any bit or bits is not. For example, given a first cache line load address is 0x1708fb1 (the 0x prefix indicates hexadecimal). Then the contents of the stream register after the load is:
Base=0x1708fb1 Mask=0x7fffff Valid=1
Subsequently, when a second cache line load address is added to this stream register, the second address is compared to the base register to determine which bits are different. The mask register is then updated so that the differing bit positions become zeros in the mask. These zeros thus indicate that the corresponding bits of the base register are “don't care”, or can be assumed to tale any value (zero or one). Therefore, these bits are no longer significant for comparisons to the stream register. For example, say the second cache line load address is 0x1708fb2. Then the contents of the stream register after this second load is:
Base=0x1708fb1 Mask=0x7fffffc Valid=1
In other words, the second address and the base register differed in the two least significant bits, causing those bits to be cleared in the mask register. At this point, the stream register indicates that the addresses 0x1708fb0, 0x1708fb1, 0x1708fb2, and 0x1708fb3 can all be in the cache because it can no longer distinguish the two least significant bits. However, it is important to note that the two addresses which have actually been loaded are considered to be in the cache. This mechanism thus guarantees that all addresses presented to the stream register will be included within it. In the limit, the mask register becomes all zeros and every possible address is included in the register and considered to be in the cache. Clearly, the mechanism described can be used to continue adding addresses to the stream register.
Every cache line load address is added to exactly one of the multiple stream registers. Therefore, the collection of stream registers represents the complete cache state. The decision of which register to update is made by the update choice logic block 704 in
In selecting a stream address register to update, the line load address is compared to all base registers combined with their corresponding mask registers in parallel. The line load address is then added to the selected stream register as described herein.
The snoop check logic block 702 determines whether a snoop address 714 could possibly be in the cache by comparing it to all of the stream registers as follows: the snoop address 714 is converted to a line address by removing the low-order bits corresponding to the offset within a cache line. This line address is compared with a single stream register by performing a bitwise logical exclusive-OR between the base register and the snoop line address, followed by a bitwise logical AND of that result and the mask register. If the final result of these two logical operations has any bits that are not zero, then the snoop address is a “miss” in the stream register and is known not to be in the cache, as far as that stream register is concerned. The same comparison is performed on all of the stream registers in parallel, and if the snoop line address misses in all of the stream registers, then the snoop address is known not to be in the cache and can be filtered out (i.e. not forwarded to the cache). Conversely, if the snoop address hits in any one of the stream registers, then it must be forwarded to the cache.
The snoop check logic 702 is duplicated for each of the N remote snoop request ports, but they all share the same set of stream registers 700.
Over time, as cache line load addresses are added to the stream registers, they become less and less accurate in terms of their knowledge of what is actually in the cache. As illustrated in the example above, every mask bit that becomes zero increases the number of cache lines that the corresponding stream registers specifies as being in the cache by a factor of two. In general, the problem of forwarding useless snoop requests to the processor (i.e., failing to filter them) becomes worse as the number of mask bits that are zero increases. Therefore, the stream register snoop filter are provided with a mechanism for recycling the registers back to the initial condition. This mechanism is based upon the observation that, in general, lines loaded into the cache replace lines that are already there. Whenever a line is replaced, it can be removed from the stream registers, since they only track which lines are in the cache. Rather than remove individual lines, the stream register snoop filter effectively batches the removals and clears the registers whenever the cache has been completely replaced. However, the new cache lines that were doing this replacement were also added into the stream registers, so the contents of those registers cannot simply be discarded.
To solve this dilemma, the stream register snoop filter performs the following: starting with an initial cache state, stream register updates occur as described previously herein. The cache wrap detection logic block 706 is provided with functionality for monitoring cache update represented by cache update signals 717 and determining when all of the cache lines present in the initial state have been overwritten with new lines, i.e. the cache has “wrapped”. At that point, contents of all of the stream registers (call them the “active” set) are copied to a second “history” set of stream registers and the stream registers in the active set are all returned to the invalid state to begin accumulating cache line load addresses anew. In addition, the state of the cache at the time of the wrap becomes the new initial state for the purpose of detecting the next cache wrap. The stream registers in the history set are never updated. However, they are treated the same as the active set by the snoop check logic 702 when deciding whether a snoop address could be in the cache. With this mechanism, the stream registers are periodically recycled as the cache is overwritten.
There are a number of ways that cache wrapping can be detected depending upon the cache update policy and the cache update signals 717. For example, if the cache specifies the line that is overwritten, then a simple scoreboard can be used to determine the first time that any particular line is overwritten and a counter can be used to determine when every line has been overwritten at least once. Any mechanism for detecting cache wrapping can be used without departing from the scope of the invention.
In a preferred embodiment, two sets of stream registers are used, but more than two sets can be used without departing from the scope of the invention. For example, in an embodiment implementing four sets of stream registers, two sets of active registers, A and B, and two sets of corresponding history registers, are implemented. In this embodiment, the A set of stream registers can contain information related to one subset of the cache, and the B set of stream registers can contain information related to a different subset of the cache. The partition of the cache into parts assigned to each set of stream registers, A and B, can be performed by dividing the cache into two equal parts, but other partitions may be used. Furthermore, the number of stream register sets can be more than two. For example, there can be one set of stream registers assigned to each cache set of a set-associative cache.
In yet another embodiment, there can be more than one history set of stream registers, allowing the active set to be recycled more frequently. However, care must be taken to manage the history registers relative to cache wrap detections so that a register is never cleared when a cache line covered by that register could still be in the cache. One way to ensure that a register is never cleared is to add history registers to the active set of stream registers and then copy all of those history registers (and the active registers) to a second set of history registers when the cache wraps. This is essentially adding a second “dimension” of history to the preferred embodiment of the stream register snoop filter as described herein.
Referring now to
Referring now to
As described herein with respect to use of the stream register snoop filter, the power of each stream register filter to block snoop requests decreases as the number of mask bits set to zero increases. For example, if all mask bits are zero, then all snoop requests must be sent through. However, supposing these mask bits were set to zero one bit at a time (i.e., each load differs from the stream register by only one bit), then, in such a case, a snoop request for an address having exactly two bits different from the stream register would be let through, even though this address cannot be in the cache. Accordingly, additional filtering capability is provided by implementing signature filters that enable detection of more complicated, or subtle, differences such as the number of different bits. The general idea is that a snoop is forwarded from a stream register only if both the mask filter and the signature filter indicate that the address might be in the cache.
Referring to
If the address 901 is a load to the cache, the signature 903 is fed to a signature register updater 904. The updater also takes the previous value of a signature register 905 as input and replaces it by a new value 906. The appropriate way to update the signature register depends on the type of signature. Let S_old denote the old value of the signature register, S_new denote the new value of the signature register, and V denote the value of the signature 903. Corresponding to the signature functions above, the signature updater 904 computes:
When a snoop request comes in, its signature is computed and compared to the signature register. It a match does not occur there, the address cannot be in the cache, so the request is filtered even if the normal stream register and mask filter indicates that it might be in the cache. A snoop is forwarded only if the signature register and mask register both indicate that the address might be in the cache.
The signature filtering mechanism is shown in
When a snoop address 1009a request comes in, it is received and processed by the mask filter 1010 producing a mask snoop request 1011. In addition, this same snoop address (shown as 1009b) and the stream register 1004 are fed to the signature function 1012 producing a signature 1013. Note that the signature functions 1005 and 1012 must be identical logic, meaning that if they have the same inputs they will produce the same outputs. The signature of the snoop request 1013 and the signature register are fed to the signature filter 1014.
This filter must determine if a request having this signature might be in the cache and its exact operation depends on the type of signature. In the case of the “scoreboard” types of signature updaters, the snoop signature is bit-wise and-ed with the signature register. If the result of this is non-zero, then a signature snoop request 1015 is made (i.e., that signal is set to 1 if a request is to be made and 0 otherwise). In the case of “maximum number of bits changed” types of signature updaters, a check is made to see if the snoop signature is less than or equal to the signature register (one comparison for each group). If all such comparisons are true, the address might be in the cache and the signature snoop request 1015 is made. The mask snoop request 1011 and the signature snoop request 1015 are AND-ed together in logic element 1016 to generate a snoop request signal 1017. If this signal is 1, a snoop request will be generated unless it is ruled out by the snoop vector lists, or an applied range filter (see
The signature register is set appropriately at the same time that the stream register is first set, or reset. For scoreboard types and max-types of signatures, the signature register is set to all zeros (indicating no bits different from the stream register).
The stream register filter relies upon knowing when the entire contents of a cache have been replaced, relative to a particular starting state—a cache wrap condition as referred to herein. A set-associative cache is considered to have wrapped when all of the sets within the cache have been replaced. Normally, some sets will be replaced earlier than others and will continue to be updated before all sets have been replaced and the cache has wrapped. Therefore, the starting point for cache wrap detection is the state of the cache sets at the time of the previous cache wrap.
In one embodiment, the cache is set-associative and uses a round-robin replacement algorithm, however other replacement implementations are possible. For instance, cache wrap detection may be achieved when the cache implements an arbitrary replacement policy, including least-recently-used and random. As referred to in the description to follow, a set-associative (SA) cache comprises some number of sets, where each set can store multiple lines (each with the same set index). The lines within a set are called “ways”. Hence, a 2-way set associative cache has two (2) lines per set. All of the ways within a set are searched simultaneously during a lookup, and only one of them is replaced during an update. Furthermore, a set can be partitioned such that a subset of the ways is assigned to each partition. For example, a 4-way SA cache may be partitioned into two 2-way SA caches. The virtual memory page table (and the translation lookaside buffer (TLB)) can provide a partition identifier that specifies which cache partition a particular memory reference is targeted at (both for lookup and update). The register that stores the way to be updated for a cache wrap needs to be big enough to store a way number. For example, 2 bits for a 4-way SA cache, or 5 bits for a 32-way SA cache. There is one such register per set because each set can wrap at a different time.
In one embodiment of the invention, the cache is partitionable into three partitions, with each partition including a contiguous subset of the cache ways, and that subset is the same within each cache set. Memory references are designated by the processor's memory management unit to be cached in one of the three partitions. Updates to a partition occur independently of the other partitions, so one partition can wrap long before the entire cache wraps. However, detecting the wrapping of a partition is identical to detecting the wrapping of the entire cache when the partition being updated is known. Thus, as referred to hereinafter, cache wrapping includes either partition wrapping or entire cache wrapping.
In order for external logic to detect cache updates, a cache must provide an indication that an update is occurring and which line is being overwritten. The logic of the preferred embodiment assumes that this information is provided by means of a set specification, a way specification and an update indicator.
a) and 19(b) depict the cache wrap detection logic of the preferred embodiment for an N-way set-associative cache. In this embodiment, it is assumed that updates to a set are always performed in round-robin order. That is, the “victim” way chosen to be overwritten is always the one following the previously-overwritten one.
a) particularly depicts one embodiment of logic implemented for detecting the wrap of a single partition of a single set (set “i” in the embodiment depicted) within the logic block 920. When this logic has detected a wrap in set i, it asserts the set_wrap(i) signal 910.
On the left-hand side of
Within the set wrap detection logic, the common partition update indicator is further qualified to act only when the update is to the particular set i associated with that logic. This is done by matching the set specifier 924 to the index of the set wrap detection logic 926.
The remainder of the logic circuits function as follows: Assume that initially, the flip-flop driving set_wrap(i) 930 is clear, indicating that the set has not wrapped, and the register 928 includes the way that must be updated to complete a set wrap. In this state, the register retains its value. When a cache update occurs, where the way 914 matches the contents of the register 928, as determined by a comparator device 919, the flip-flop driving set wrap(i) 930 is loaded with logic 1, causing set_wrap(i) 910 to assert. Thereafter, cache updates cause the updated way 914 to be stored in the register 928, so the register 928 effectively tracks those updates. When all cache sets have wrapped, the combined cache_wrap 912 signal is asserted as shown in
It is thus understood that there is one register per set that stores the number of a way and when that way is overwritten, then the set has wrapped. However, the sets wrap at different times (depending on the access pattern), and the entire cache is not considered to have wrapped until all sets have wrapped. At that point, the state of the victim way pointers (i.e. pointer to the last way that was overwritten; one per set) becomes the new initial condition for detecting the next cache wrap. The first embodiment accommodates this requirement by having the register described above keep track of ways that are overwritten between the time that it has wrapped and the time that the entire cache has wrapped. Then when the whole cache wraps, it stops tracking the overwritten ways and becomes the basis for comparison for determining when the set wraps again.
In a second embodiment of the cache wrap detection logic, a counter is implemented, so when the whole cache wraps, all set counters are reset to the number of ways in the partition. As ways are overwritten, the counters count down; and when a counter reaches zero, then the corresponding set has wrapped. When all counters reach zero, then the cache has wrapped and the process starts again.
According to this second embodiment, the set wrapped detection logic provided within the box 920 depicted in
A third embodiment of the cache wrap detection logic, shown in
While the preferred embodiment of the present invention is practiced in conjunction with a write-through cache, wherein snooping only occurs on write requests, and the results of a snoop action are the invalidation of a local data copy, the invention is not so limited. For instance, the invention can also be practiced in conjunction with write-back cache organizations. In accordance with a write-back cache, a coherence protocol will include additional transactions, e.g., including but not limited to, those in accordance with the well-known MESI protocol, or other coherence protocols. In accordance with a coherence protocol for writeback caches, read transaction on remote processors cause snoop actions to determine if remote caches have the most recent data copy in relation to the main memory. If this is the case, a data transfer is performed using one of several ways, including but not limited to, causing the processor having the most recent data to write the data to main memory, directly transferring the data from the owner of the most recent copy to the requestor, or any other method for transferring data in accordance with a snoop intervention of a specific protocol. In accordance with this invention, a snoop filtering action can be used to determine an accelerated snoop response.
While the preferred embodiments have been described in terms of fixed interconnection topologies, and fixed snoop filtering operations, in one aspect of the present invention the snoop filtering subsystem has programmable aspects at one, or more, levels of the snoop filter hierarchy. In accordance with one embodiment of a programmable feature of the present invention, the interconnect topology is selected. In accordance with one variety of programmable topology, the one-to-one and one-to-many relationship between different filters in a topology is selectable. In accordance with another aspect of a programmable embodiment, the order in which a first snoop filter, and then a second snoop filter is accessed, or alternatively, a first or second snoop filter are accessed in parallel, is configurable under program control.
In accordance with yet another aspect of yet another embodiment of a programmable feature of the present invention, the operation of a filter subunit is programmable. This can be in the form of configurable aspects of a snoop filter, e.g., by configuring programmable aspects such as associativity of the cache being snooped, the coherence architecture being implemented, and so forth. In another aspect of a programmable filter subunit, the filter subunit is implemented in programmable microcode, whereby a programmable engine executes a sequence of instructions to implement the aspects of one or more preferred embodiments described herein. In one aspect, this is a general microcode engine. In another aspect, this is an optimized programmable microcode engine, the programmable microcode engine having specialized supporting logic to detect snoop filter-specific conditions, and, optionally, specialized operations, such as “branch on cache wrap condition”, specialized notification events, e.g., in the form of microcode engine-specific exceptions being delivered to the microcode engine, such as “interrupt on cache wrap condition”, and so forth.
In yet another embodiment of a programmable feature of the present invention, parts or all of the aspects of snoop filtering are implemented incorporating a programmable switch matrix, or a programmable gate array fabric. In one of these aspects, the routing between snoop subunits is performed by configuring the programmable switch matrix. In another aspect of this programmable embodiment, the actions of the snoop filter unit are implemented by configuring a programmable gate array logic block. In another aspect of the present invention, the entire snoop filter block is implemented by configuring at least one field-programmable gate array cell.
In accordance with another embodiment of a programmable feature of the present embodiments, one of more snoop filter subsystems can be disabled, certain snoop filtering steps can be bypassed, or snoop filtering can be disabled altogether. In one embodiment, this is achieved by writing the configuration of the snoop filter in a configuration register. In another embodiment, this configuration can be selected by input signals.
While there has been shown and described what is considered to be preferred embodiments of the invention, it will, of course, be understood that various modifications and changes in form or detail could readily be made without departing from the spirit of the invention. It is therefore intended that the invention be not limited to the exact forms described and illustrated, but should be constructed to cover all modifications that may fall within the scope of the appended claims.
This application is a continuation application of U.S. Ser. No. 11/093,127, filed Mar. 29, 2005, which relates to commonly-owned, co-pending U.S. patent application Ser. Nos. 11/093,130; 11/093,131; 11/093,152; 11/093,132; 11/093,160; 11/093,154 all filed on even date herewith and incorporated by reference as if fully set forth herein.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 11093127 | Mar 2005 | US |
| Child | 12126674 | US |
