The invention relates to a parallel hardware hypervisor system for virtualizing application-specific supercomputers.
Maturity of processor architecture research: The general-purpose processor architecture research field has matured, with attempts to further increase the performance of general-purpose processors presently encountering (i) frequency, (ii) power, (iii) design complexity, and (iv) memory wall barriers. However, the need for increased performance and reduced power continues to exist.
Difficulty of parallel programming: Abandoning the extremely convenient, easy-to-use sequential programming model and programming explicitly for parallel processors constitute one way for increasing performance. Recent multi-core processor architectures [5] that are enabled by increasing VLSI densities indeed encourage this approach. However, programming a parallel multi-core processor system is not a natural and easy task, due to, e.g., race conditions, deadlocks, and non-deterministic bugs that are hard to track. Increased parallelism in general-purpose processors has in fact increased the difficulty of programming and using them [2].
Inefficiencies of the hypervisor and the operating system: Sharing of computing resources among different independent applications and virtual machines has been emphasized at least since the days of early mainframes [1]. This emphasis on resource sharing continues to this day. Recently, Cloud Computing [3] and Virtualization [4] have emerged as preferred methods of offering computing and application services with resource sharing. By breaking the barriers of the traditional in-house IT shop approach, cloud computing offers centralized high performance computing resources, economies of scale, and radically higher degrees of efficiency. For example, a large cloud computing data center, along with a fast and reliable encrypted network, can greatly amplify the performance of an inexpensive client device, while preserving the security properties of an in-house IT shop.
However, cloud computing today relies on operating systems or hypervisors that are designed in software, and that lack scalability. For example, the cost of an interrupt may involve substantial overhead (e.g., ten thousand instructions) in today's operating systems. Moreover, the transition between privilege levels (as in an interrupt or system call) requires a global serialization/pipeline flush in general-purpose processors. The schedulers within operating systems and hypervisors alike are not designed in an algorithmically parallel scalable way, to handle massively parallel systems. At the extreme performance levels that will be needed in the future, such serialization overheads will become important. To alleviate the severe performance slowdown consequences of Amdahl's law, the slowdown effects due to both the OS and the hypervisor must be reduced.
Prevailing solutions: Current computer industry focus areas include two prevailing approaches, namely: energy-efficient multi-core processors [5] and hybrid computing architectures [6], which, while not directly addressing the significant problems mentioned above (namely, the difficulty of parallel programming, and the inefficiency of the OS and hypervisor), do promise to increase performance and to reduce power. We will review the hybrid computing architectures, since they are most relevant to application-specific supercomputers, the subject of the present document.
In general-purpose hybrid computing architectures, the acceleration unit consists of graphics processing units (GPUs) with their own specialized Instruction Set Architecture [6]. These acceleration units are capable of accelerating graphics applications, as well as a range of additional high performance computing applications, provided that suitable parts of the applications are re-coded to expose explicit parallelism and to take advantage of the massively parallel architecture of specialized processors.
By contrast, reconfigurable hybrid computing architectures (reconfigurable computers) deploy field programmable gate arrays (FPGAs) as the acceleration unit, and offer more flexibility. Typically, a collection of one or more FPGAs acts as a co-processor to each general-purpose host processor [7] [8]. While arbitrary code in general cannot take advantage of the FPGAs using today's tools, suitable code fragments can again be recoded to expose explicit parallelism and then compiled with a high-level tool to run on the FPGAs.
Even though the commercial systems with FPGAs are very promising in boosting the application performance with less power than traditional servers, they suffer from a few shortcomings:
Our approach: The present document's system does address the two significant problems (difficulty of parallel programming, inefficiency of the OS and hypervisor) mentioned above. It also distinguishes itself from the cited art in at least the following ways:
We describe a parallel hypervisor system for virtualizing application-specific supercomputers, where the system comprises:
A virtual or physical tile contains arbitrary digital circuits. The hypervisor system can be used to implement cloud computing with software applications accelerated by application-specific virtual supercomputers. Physical hardware resources can be incrementally increased or decreased on-demand for each application, at the physical tile granularity. Features of the hypervisor system include:
The hypervisor design avoids system-wide serialization points, through the parallel handling of cache misses and coherence actions within the local virtual tile to physical tile caches described above, by using the following key hardware units:
Multiple owner units: the set of all virtual tiles in the system is partitioned, and one owner unit is assigned to each partition. An owner unit maintains the map from each virtual tile in its partition to a physical tile (if the virtual tile is mapped) or to NULL (if the virtual tile is not mapped). Using multiple owner units simultaneously allows parallel, independent search and tile pre-emption activities.
A monitor unit continuously obtains statistics about activity in the system. It then analyzes the statistics and provides replies to requesting owner units, in a parallel manner, to suggest a new physical tile to pre-empt to each owner unit, according to a tile replacement policy.
We will now describe the details of a parallel hypervisor system for virtualizing application-specific supercomputers, where the system comprises:
The preferred embodiment of the hypervisor consists of the following major parts:
We will first describe the overall hardware structure of the preferred embodiment of the hypervisor system. The hypervisor system is organized in hierarchical enclosures, very much like a non-virtual (real) supercomputer. It comprises the following, starting from the leaves of the hierarchy and going towards the root:
“Incomplete hypercube” is used in the sense that the total number of chips in the system need not be a power of two. The total number of chips in the system can be any number greater than or equal to one. Although we will stick to the incomplete hypercube topology in this document, for systems with a very large number of chips, a cube-connected cycles topology (where a communicating group of chips serves a single hypercube node, therefore effectively increasing the number of hypercube links of each node) can be used.
An Example Illustrating the Operation of the Hypervisor
To motivate the forthcoming detailed hardware description of the hypervisor system, we will start by describing the operation of the hypervisor system on a small example.
When we say “cache” within the following text, we do not mean a data cache or instruction cache. The caches of the hypervisor system implement a mapping from virtual tiles to physical tiles; they do not contain data. These caches help speed up the transmission of messages within a virtual supercomputer.
The key to fast sending of messages within the virtual supercomputer is a set of globally coherent first level caches mapping virtual tiles to physical tiles, such that there is a cache present right next to each physical tile. Such a local cache allows:
The virtual tile numbered −1 is special within each virtual supercomputer: it is used as a message exchange gateway to the corresponding software application running on the host processor system, which is reached via a PCI Express connection. This virtual tile number −1 is permanently pinned to a physical tile during the lifetime of the application, for simplifying message routing.
Each (application, virtual supercomputer) pair in the hypervisor system is assigned a unique application id number (e.g., 0x00=application A instance 0, 0x10=application B instance 0, 0x11=application B instance 1, . . . ).
As opposed to local virtual tile numbers, which are integers in the range −1, 0, . . . , maximum virtual tile number within the given virtual supercomputer, a global virtual tile uniquely identifies any virtual tile within any virtual supercomputer in the hypervisor system, and is a pair (application id, local virtual tile number within this application). In the following text, a virtual tile (when not explicitly specified as a “local virtual tile” or “global virtual tile”) will mean a global virtual tile.
The owner unit for a virtual tile is found by computing a simple hash function of the virtual tile.
Notice that at this point, virtual tile A1 has not been used and has not been allocated in a physical tile. If the virtual tile A1 is not needed by the current inputs of the computation the virtual supercomputer for application A is engaged in (i.e., A0 is sufficient for accelerating application A for its current input), then no message will be sent to A1 and no physical tile will be allocated for A1. The present hypervisor system therefore allows on-demand increase of hardware resources for a given virtual supercomputer, delaying the allocation of a physical tile until (if ever) it is actually needed. Similarly, we will see that physical tiles that have remained idle will be pre-empted, resulting in an incremental decrease of resources for a given virtual supercomputer. Incremental increase and decrease of resources is an essential requirement of ordinary software cloud computing; the present hardware hypervisor system provides this feature for incremental provisioning of hardware acceleration resources to applications.
Notice that the mapping (A0→P2) has remained in the local cache of P3. If it remains unused, it will eventually be evicted from this local cache.
In the illustrative small example above, we described the effects of message transmissions on the hypervisor system, as if the each message were transmitted following a global sequential order on the message transmissions. In reality, when two messages are independent, they will be transmitted in any order or in parallel, and the tile pre-emption and cache coherence actions will also occur in parallel, thanks to the multiple owner units. We will describe the highly parallel hardware implementation of the hypervisor system in detail and also show how race condition errors are avoided, in the sections below.
Primitive Hardware Building Blocks of the Hypervisor
At this point we incorporate by reference the co-pending, co-owned non-provisional U.S. patent application Ser. No. 13/296,232, entitled “Method and system for converting a single-threaded software program into an application-specific supercomputer”. This patent application will be called [Supercomputer] from this point on.
The present document describes a hypervisor system comprising virtual supercomputers, while [Supercomputer] describes a method and system for creating non-virtual (real) supercomputers; therefore, the present document's subject is different. However, referring to [Supercomputer] (i) clarifies and shortens the present hypervisor system's baseline hardware description, and also (ii) provides an important kind of preferred embodiment.
The present hypervisor system will work with any physical tiles, virtual tiles and virtual supercomputers wherein:
However, when the technology described in [Supercomputer] is combined with the present hypervisor, we obtain an important kind of specialized physical tiles, virtual tiles, and virtual supercomputers such that:
In an attempt to make the present document more self-contained, we will now also briefly summarize the features from [Supercomputer] which are re-used in the present hypervisor system. These features are specifically helpful for implementing the components of the present hypervisor at a low hardware level; although an experienced hardware designer may choose other techniques which will work equally well for the same implementation.
Starting from the flat non-partitioned design of the hypervisor, to be described in detail below, the design partitioning and chip unioning technique described at least in the paragraphs [00169]-[00190] and Appendix K of [Supercomputer] will also be used to create the “union chip” of the hardware hypervisor design, a chip that is capable of realizing any partition of the partitioned hypervisor design. This union chip will be called the cloud building block chip. The entire cloud data center will consist of copies of the cloud building block chip in hierarchical enclosures (such as rack modules and racks), wherein the copies of the cloud building block chip will possibly differ only in the types of physical tiles contained in it.
An example of design partitioning and chip unioning, as it relates to creating a scalable hypervisor, will be given in the paragraphs below beginning with the words “Let us summarize the design partitioning and chip unioning technique”.
Key Hardware Components of the Hypervisor
The Networks and Components
Armed with the basic hardware building blocks of [Supercomputer], we will now describe the hardware hypervisor system in detail.
Again referring to
Of course, the flat design described above and in
A union chip, such as the cloud building block chip in
Let us summarize the design partitioning and chip unioning technique of [Supercomputer], and show with an example how a collection of union chips work together to realize the function of a flat non-partitioned design of a cloud computing data center: Assume we are given, for the sake of an example, an original flat design of a small data center, with 14 physical tiles (including 2 application placeholder tiles with PciE connections), 3 owners, and one monitor. Since only a small number of components will fit in each chip, after design partitioning and chip unioning, assume that this flat design leads to 4 cloud building block chips 0, 1, 2, and 3, each with:
As an example of message routing using the cloud building block chips, here is how owner unit 2 in chip 2, sends a tile_request (“give me a physical tile to pre-empt”) message over the replacement tile selection network, to the monitor unit in chip 0. Gray code versions of chip numbers 0, 1, 2, and 3 are used (0=Gray 00 1=Gray 01 2=Gray 11 3=Gray 10), since Gray codes are more convenient for the purpose of deterministic hypercube routing, A scalable hypercube network connects the cloud building block chips.
The net result is that the collection of the 4 cloud building block chips displays behavior identical to the original flat design with 14 physical tiles, 3 owners, 1 monitor and 2 PCI Express connections, as described above. More details of the design partitioning and chip unioning algorithms are described in [Supercomputer].
The “hypervisor storage” for keeping the saved state of virtual tiles is realized with off-chip DRAM units. For saving and restoring tile states, it suffices for each owner unit to access the local DRAM unit closest to the chip containing the owner unit. We have not included the DRAM DDRn controllers in the figures, for simplifying the figures, although each cloud building block chip will have at least one such controller. The DDRn controllers in each chip are connected to DRAM units packaged in, e.g., DIMMs (Dual Inline Memory Modules) on the board. The physical tiles implementing the virtual supercomputers may also share different regions of the same local DRAM resources on the board.
We have used many different networks for different functions in this design as a “separation of concerns” simplification, as in software design practice. Of course, the number of networks can be reduced by resource sharing, for example, by using (optimization 5, starting on p. 145 of [Supercomputer]) repeatedly, or by creating virtual networks each with their separate input and output FIFOs, where the virtual networks are implemented on a single common physical network or bus connecting the hardware components.
Message Formats within the Hypervisor System
In this section we will describe the message formats used in the hypervisor system. We will start with an example of a message:
The message format first indicates whether the message is being sent or received by the present hardware unit, and whether the message is a request or a response, and further identifies the network where the message is sent or received. The remaining part of the message is sequence of field specifications of the form: Field=(explanation=variable) when the value of the variable is used for creating the message field or for checking the value of the message against an expected value, or of the form Field=(variable=explanation) when the message field (which already has a value in this case) is assigned to the variable. “me” appearing in messages identifies the number of the current hardware unit sending or receiving the message. Variable names have local scope within their message exchange description, unless otherwise specified. The Field=(explanation) form is also used, in case no variables are needed. We will explain the message fields below:
Note that ceil(log2(number of possible values of the field)) bits are needed to encode a field in a message, within the context of the messages described in the present document. In particular, when there is only one possible value for a field, 0 bits are required to encode the field, and therefore such a field will not physically appear in the message. For example, the Opcode field in the single operation code case will not be physically present in the messages; in this case, the Opcode is provided only as a convenience for the reader.
Long messages will be broken up into a sequence of words with an end-of-data bit=0, ending with a word whose end-of-data bit=1, as in [Supercomputer]; this variable length encoding does not change the meaning of the message. The order of the fields within a message, and the particular binary values for representing constant fields, are not important, as long as a precise contract for the message format is followed throughout the design.
The Reconfigurable Physical Tile
The physical tile is a hardware component that can be reconfigured to implement one or more virtual tiles of one or more virtual supercomputers within the hypervisor system.
Referring to
The physical tile's internal operation is defined mainly by the virtual tile it is currently configured to implement. The hypervisor system does not need to understand the internal operation of the virtual tile currently being implemented, as long as the virtual tile complies with the requirements described in the paragraph above beginning with the words “The present hypervisor system will work with”. In this section, we will specify the following key behavior of a physical tile relevant to the operation of the hypervisor system:
The physical tile's slave port of the pre-control network facing the physical tile harness accepts the following requests and sends back the following responses:
At system reset time, each physical tile is in the stopped condition, except for application placeholder tiles (described in the section below entitled “Application placeholder physical tiles”) which are running and are ready for any message exchanges with their respective software applications.
Pre-conditions of shutdown_and_read_state: When the shutdown_and_read_state request is received by a physical tile p currently running virtual tile v, all local first level cache entries (v→p) within physical tile harnesses in the hypervisor system, pointing to the present physical tile p, should be already invalidated, and any remaining incoming messages should have been received by the physical tile harness attached to the present physical tile. Notice that, as it will be seen in the owner unit section below, once a local cache entry (v→p) is invalidated, further messages to virtual tile v will be blocked until the owner unit of v completes the de-allocation of (v→p) and starts processing the waiting access_request commands for v, to reallocate v in possibly a new physical tile.
Post-conditions of shutdown_and_read_state: Notice that after processing a shutdown_and_read_state request: all pending incoming messages to virtual tile v on physical tile p are consumed (no incoming messages are left in the network); and further incoming messages to v are blocked in the owner of v. But the outgoing messages sent by the virtual tile v on physical tile p may remain undelivered to their destinations; these outgoing messages may sit in the network for an arbitrary time. These outgoing messages should be delivered to their destination virtual tiles, before the virtual tile v is reallocated to a different physical tile p′≠p, to prevent re-ordering of messages coming out from virtual tile v. Note that the networks of the preferred embodiment use deterministic routing and guarantee that the messages from one given input port to one given output port will not be reordered in the network; but there is no guarantee regarding the order of delivery of messages from different input ports. Hence, when a virtual tile is migrated to a new physical tile, a drain command is required, to ensure the delivery of the pending undelivered messages that emanated from the same virtual tile, while it was allocated to its prior physical tile.
We already provided above a method for shutting down a virtual tile v (currently on physical tile p) which frequently reads its input, by simply honoring shutdown requests only when the virtual tile's normal inbound pre-communication input FIFO is empty. For shutting down a virtual tile v (currently on physical tile p) which rarely reads its input, we propose another method:
A physical tile's unidirectional master port of the statistics network facing the monitor, periodically issues the following requests:
A physical tile's unidirectional master port of the outbound pre-communication network, and unidirectional slave port of the inbound pre-communication network, where both networks face the attached physical tile, accomplish the inter-virtual-tile communications within the application-specific virtual supercomputer. Both outgoing and incoming messages have the same message format:
Each virtual tile implemented on a reconfigurable physical tile is made to believe (with the help of the hypervisor infrastructure) that it is communicating natively with other virtual tiles of the same virtual supercomputer, as if the supercomputer were implemented with native, non-virtual hardware. In reality virtual tiles are allocated to physical tiles on demand, and then possibly pre-empted (de-allocated from the physical tile), for example, after the virtual tile has remained idle for a sufficiently long time. The virtual destination tile within such a message is needed, for looking up the corresponding physical destination tile number. The virtual source tile within such a message is also needed, for keeping track of any undelivered messages emanating from this virtual source tile. Therefore, a pre-communication network message for inter-virtual-tile communication, should meet the following pre-communication message format requirement:
Assuming the local virtual tiles of a given virtual supercomputer performing the real hardware functions are numbered 0, 1, 2, . . . , n−1, it is convenient to create a new local virtual tile of the same virtual supercomputer, numbered −1, whose only job is to relay messages to and from the software application, which this virtual supercomputer accelerates. This way, messages exchanged between the local virtual tiles and the software application do not need to be treated as a special case with respect to message routing. Given a hypervisor system implementing m (software application, virtual supercomputer) pairs, using m general purpose commodity host processors each running a software application that is accelerated by a virtual supercomputer, we can create m application placeholder virtual tiles, and permanently map them to fixed physical application placeholder tiles within the hypervisor, that will not be de-allocated. Each application placeholder physical tile will communicate point-to-point with a PCI Express external communication device that leads to the correct host processor running the corresponding software application. Thus, when a host application sends a message to a local virtual tile of its virtual supercomputer, this message enters the hypervisor system at the dedicated PCI Express connection and application placeholder physical tile tied to this host processor. A message sent by the software application will appear to be coming from local virtual tile −1. When a local virtual tile numbered 0, 1, . . . of the virtual supercomputer wishes to send a message to its software application, it will send the message to its application placeholder local virtual tile −1, which will in turn forward the message to the software application over the PCI Express connection.
Physical Tile Harness Unit
Referring to the schematic in
Internal memory state: The internal memories and registers of a physical tile harness are:
As it will be described below, the control FSM and the outbound communication request FSM of the physical tile harness both share (i) the local virtual tile to physical tile cache called L1 (ii) the lockedDest and lockedSource sets of virtual tiles. The outbound communication request FSM, the outbound communication response FSM, as well as the control FSM share the outstandingByDest and outstandingBySource counter arrays. The accesses to these shared data structures must be made to appear atomic, which can be achieved by a multi-ported design and/or a network for arbitration.
The physical tile harness includes the following internal finite state machines (FSMs), which will each be described separately.
Outbound Communication Request FSM:
Outbound Communication Response FSM:
The outbound communication request FSM has the following FIFO interfaces
The Outbound Communication Response FSM has the Following FIFO Interfaces
The “Outbound communication request FSM” performs the following steps repeatedly, in an unending loop:
Notice that while an abandon request is in progress, concurrent invalidate requests can make the L1 cache set smaller automatically, and therefore the need to abandon an entry may go away by itself
The transactional implementation of abandon described above is required because, depending on the network contention, there may be many additional transactions in the owner regarding v′ (such as deallocating v′ from p′, allocating v′ on a different physical tile p″), while the abandon message for (v′→p′) is in transit from the present physical tile to the owner of v′
The “outbound communication request FSM”, written in sequential code here, can be “software pipelined” (i.e., iterations n+1, n+2, . . . can be started before iteration n is finished) by correctly respecting dependences. For example, when there are two back to back access_request commands to the same virtual tile, but the first one misses in the local first level cache, the second request must wait for the first one to update the local first level cache. However, two back-to-back access_request commands to different virtual tiles can proceed in parallel/pipelined fashion. Messages from one given virtual tile to another given virtual tile should never be re-ordered, since not reordering messages between a pair of virtual tiles is a guarantee the hypervisor gives to all virtual supercomputers.
The “outbound communication response FSM” can also be software pipelined, so that a deeply pipelined implementation of the outstandingBySource and outstandingByDest data structures can be utilized.
Inbound Communication FSM:
This FSM has the following FIFO interfaces
The inbound communication FSM executes the following steps in an unending loop:
This FSM has the following ports:
The control FSM executes the following steps in an unending loop:
There are no other kinds of control requests.
The control FSM will not be software pipelined within the physical tile harness, since physical tile configuration operations cannot be easily software pipelined. But an owner unit can overlap control network requests to different physical tile harnesses when dependences permit.
Owner Unit
Referring to the schematic in
Internal Memory State:
The owner unit has the following internal FSMs:
The Lookup FSM
The lookup FSM has the following ports
The lookup FSM executes the following steps in an unending loop:
The lookup FSM waits for a request from the slave port of the lookup network facing physical tile harnesses.
Access Request
If the incoming message is an access request of the form:
If the request is an abandon request of the form:
The lookup FSM can be software-pipelined subject to normal sequential execution constraints. For example, if a first access_request to a virtual tile results in a miss in the ptile map, a second access_request to the same virtual tile must wait until the first request is processed and the ptile data structure is updated. However, a second request to a different virtual tile can proceed independently of the first request.
Allocation/Deallocation FSM
The allocation/deallocation FSM has the following ports:
The allocation/deallocation FSM shares the ptile, priorPtile and the sharers data structures with the lookup FSM. The accesses to these data structures should be atomic.
The allocation/deallocation FSM performs the following steps in an unending loop
Deallocate Request
If there is a deallocate request of the form:
The receiving and writing of the tile state data should be pipelined, for large states.
Otherwise, if there is an allocate request of the form:
If there is a drain request of the form:
The allocation/deallocation FSM can be software pipelined, subject to sequential dependencies.
Monitor Unit
Referring to the schematic in
The monitor unit is used to detect the activity within each of the physical tiles, analyze the activity and suggest the best physical tile to pre-empt to owners who request a new physical tile to pre-empt. Each physical tile periodically sends its state to the monitor unit. In a system with N physical tiles, this can be done by an ordinary N to 1 incomplete butterfly sub-network as described in [Supercomputer], which can also cross chips in the usual way. But creating a customized pipelined token-ring network to achieve the N to 1 unidirectional communication requires less hardware. The customized pipelined token ring network can be implemented by a 1D torus (or ring) network which also passes through the monitor unit. Immediately after system reset time, for each physical tile p in the system, a packet that shall be owned and updated by p is injected into the ring, initially indicating that this tile p is nor working (i.e., idle). Normally, each physical tile forwards each incoming packet to the next node in the ring. However, when the physical tile's own packet (a packet whose id field is equal to the present physical tile) is passing by, the packet is updated with the present physical tile's current status, before being forwarded to the next node in the ring. The monitor is located between the last physical tile and the first physical tile in the ring. The monitor unit gets a packet from the last physical tile, updates its data structures as the packet from each physical tile passes by, and forwards the packet to the first physical tile. When asked for a tile to pre-empt, the monitor unit analyzes the data from all the physical tiles and returns the best tile to pre-empt, according to its replacement algorithm.
A simple scalable implementation of the Least Recently Used policy: We begin with a scalable baseline algorithm for true LRU replacement of tiles. Let us call the time from the point where a physical tile's own packet passes the physical tile and the point where its own packet passes the physical tile again, a time interval of the physical tile. Assuming that each time interval where the physical tile was active at least for one cycle, is considered a “reference” to the physical tile (as if in a reference to a data page in a virtual memory system), the least recently used algorithm can be simply implemented by mimicking the following software algorithm for LRU insertion in a doubly-linked list, as shown in the code below. Two sentinel list elements called “back” and “front” are placed at the back and front of a doubly linked list. A “reference” to a physical tile i consists of a deletion of node i from its current location (loads from flink[i] and blink[i], and stores into flink[blink[i]] and blink[flink[i]]) and a re-insertion of physical tile i just before front sentinel element (stores into flink[blink[front]], where blink[front] is cached in a register, and into flink[i] and blink[i]). The number of loads/stores is as follows: 1 load from the flunk array, 1 load from blink array, 2 stores into the blink array, and 3 stores into the flunk array. The 2 loads can be done in parallel in step 1, and then the 5 stores can be done in parallel in step 2, if memory port resources permit. Depending on the number of ports in the available memory arrays and the total number of tiles, the entire “reference” operation will require only a few cycles. The number of ports of the memory arrays can be increased in known ways, e.g., by bank-interleaving and/or by using multi-ported arrays.
The Internal Data Structures
The internal data structures of the monitor unit are as follows
At initialization time: the LRU doubly linked list is initialized to a default order, e.g., the sequential order of the physical tiles as shown above. For all physical tiles p, vtile[p] is set to NULL and working[p] is set to false, and isfree[p] is set to true. But, for pinned physical tiles p representing the application placeholder tiles, vtile[p] is set to virtual tile −1 of the respective application and working[p] is true, and isfree[p] is set to false (so that p will never be pre-empted).
The Monitor FSM
The monitor FSM has the following ports:
The monitor FSM repeatedly performs the following in an unending loop:
The monitor should answer requests sufficiently faster than the average tile replacement rate in the entire hypervisor system. Otherwise, the monitor will become a bottleneck in the hypervisor system. The optimization described in the section below entitled “4. Alternative physical tile replacement algorithms for the monitor unit” describes ways to accomplish this scalability requirement.
Solutions to Race Conditions
In this section, we will summarize five important potential race condition errors within a highly parallel implementation of a hypervisor, and show how these errors are eliminated by the present design. These race conditions will also help explain the design choices made in the present preferred embodiment.
Access Request Followed by Invalidate Causes Invalidate to be Lost
Solution: the “invalidate v” request from o1 to p1 will find v locked in p1 (by virtue of the lockedDest data structure of a physical tile harness which is checked by invalidation requests). o1 will get a negative acknowledgement for the invalidation request. The failing invalidation request will then be retried by o1.
Superfluous Abandon
Since the local cache is not instantly updated after changes to the owner data structures because of network delays, an abandon request for (v→p) can potentially be sent out by p1 and can then spend a lot of time in the network, even though the (v→p) mapping has already been deleted at the owner of v, p1 has been removed as a sharer of this mapping, and further changes have been done for v at the owner, during the transit time of the abandon message. Here an sequence of events showing the incorrect race condition:
Solution: abandon is made transactional; it is either committed or aborted. If o1 does not still have v mapped top or p1 is not a sharer of the (v→p) mapping, abandon v will get a negative acknowledgement and the abandon request will become a no-op. Another abandon (possibly to a different virtual tile) can be retried by p1, if needed for making space in the local cache of p1.
Incoming Message Destined to Virtual Tile v Arrives Late, after v has been Deallocated
Obviously, we do not want a message going to virtual tile v to arrive at a physical tile p, after the destination virtual tile v has been deallocated from p. This is solved by ensuring, with extra quiescence detection hardware (outstandingByDest outstanding message counter array and an additional reverse subnetwork where acknowledgements flow in the reverse direction of the regular communication messages), that all pending messages going to v at p have arrived at v at p, before v gets deallocated from p.
Incorrect Message Reordering Due to Migrating a Virtual Tile
Here is a sequence of events demonstrating an incorrect message reordering
Solution: With extra quiescence detection hardware (outstandingBySource counters, acknowledgement paths in communication network), messages from v1 on p1 are drained from the network, i.e., messages are made to reach their destination before v1 is reallocated on a different physical tile. In case draining the messages from v1 on p1 is not possible (because for example, of a circular wait/deadlock condition), v1 is again allocated to its old physical tile p1 without draining its old pending messages, in which case message reordering will not occur.
A circular wait/deadlock condition can occur when attempting to drain messages, for example, when an access request no. 2 for a message from v1, is waiting in the same owner's input FIFO for an access request no. 1 for a message to v1, where v1 is currently not allocated in any physical tile. We have chosen the present simple way to solve this deadlock problem (reallocate v1 in its old physical tile if unable to drain its pending outgoing messages). Reordering the access requests in the owner access request queue may be another way to avoid this kind of deadlock.
Physical Tile Gets Preempted for a Second Time while a Logically Earlier Preemption is in Progress
Here is a sequence of events demonstrating a preemption race condition:
Solution: at the time the physical tile p1 is returned by the monitor to o1, the physical tile p1 becomes locked. It will be unlocked only when o1 has finished all reconfiguration activities and sends a “tile_unlock” request for this physical tile p1 to the monitor. When all eligible tiles are locked, the monitor returns a negative acknowledge to tile requests, so the request will be retried.
Without locking, repeated choice of the same physical tile by the monitor is quite possible, for example, when the eligible physical tiles satisfying a tile request are few in number.
Optimizations
Apart from the baseline hypervisor described above, various optimizations of a hypervisor are possible. We list these optimizations and additional features below.
1. Obtaining a Virtual Supercomputer Automatically from a Single-Threaded Software Application
This optimization is facilitated because a method to obtain a non-virtual (real) supercomputer from a single-threaded software application is already described in the co-pending, co-owned US patent application [Supercomputer], which has already been incorporated by reference herein, around the paragraph above beginning with the words “At this point we incorporate by reference”. Here, we will provide the enhancements to [Supercomputer] in order to:
Much of the technology described in [Supercomputer] can be used verbatim in the present hypervisor system, once a one-to-one correspondence between the concepts of [Supercomputer] and the concepts of the present hypervisor system is established. Here is the required one-to-one correspondence:
The union chip hardware produced by the method of [Supercomputer] is adapted with slight modifications for use as a physical tile of the present hypervisor system, as follows:
For construction of a physical tile, the chip unioning technique described in paragraphs [00169]-[00190] and Appendix K of [Supercomputer], is used. The union chips of [Supercomputer] are configured once during system initialization time, by initializing the configuration memory in each chip. The configuration memory identifies the particular partition which will be realized by the union chip. But in the hypervisor system, each physical tile (union chip of [Supercomputer]) will be reconfigured multiple times, on demand. Therefore, not only the configuration memory but also the normal (execution state) memory and registers of the physical tile need to be read out and written. Assuming the simple method of stopping the clock is used to stop a virtual tile, reconfiguring a physical tile will consist of writing the configuration memory as well as the normal execution memory and registers, through one or more scan chains. Circuitry will be added to stop the clock, read out the registers and memories (including the configuration memory), write the registers and memories (including the configuration memory), and finally restart the clock.
A union chip of [Supercomputer] with n hypercube links will support a real supercomputer system having (2n-1)+1 to 2n chips, and will also include an incomplete hypercube deterministic router within it. But for the physical tile of the hypervisor, the partitioned communication network among physical tiles will already have such incomplete hypercube deterministic routing; therefore, it is not necessary to have n links, nor is it necessary to do internal hypercube routing within the physical tile. The physical tile will thus be simplified, and its internal I/O controller will have only one external communication I/O link (a sending FIFO interface (outbound pre-communication) and a receiving FIFO interface (inbound pre-communication)), as if it were part of only a 1-cube.
Based on the techniques described in detail in the specification and claims of [Supercomputer] and also in the present section, here are then the steps of a method for automatically converting an arbitrary single-threaded software application to the pair (modified version of said application, virtual supercomputer that accelerates the said application), which can then be integrated and deployed within a hypervisor system. The first four steps below are taken directly from [Supercomputer]. In the method below, the term “union chip” used in [Supercomputer] has been changed below to “union module,” since now it is not by itself a chip; it instead has become a physical tile that is part of a larger chip (the cloud building block).
2. Semi-Reconfigurable ASIC Physical Tiles
In our preferred embodiment of the hypervisor system, multiple versions of the physical tiles can be created in ASIC technology, each one customized for an important customer application. Also, another physical tile version in ASIC technology can realize a virtual tile of the “compiler-friendly general purpose supercomputer” (as described in at least the optimization 5 starting on p. 144, paragraphs [00274]-[00275] and FIGS. 60 and 61 of [Supercomputer]). Then, the physical tiles can be distributed within the data center, based on the expected percentage of computational resource usage by customer applications. For example, a cloud building block chip containing the physical tiles of a given less frequently used application A can be assigned to only one rack module, while cloud building block chips containing the physical tiles of a more frequently used application B can be placed in an entire rack. The remaining rack modules can consist of physical tiles implementing the “compiler-friendly general-purpose supercomputer union chip” [Supercomputer]. Physical tiles built out of FPGA technology can also be included in the data center, for new virtual supercomputer development (before an ASIC version of the physical tile is released). Either “compiler friendly general purpose supercomputer union chip” [Supercomputer] physical tiles, or FPGA physical tiles can be used for applications for which there is no ASIC physical tile yet, or for applications which do not warrant the development of an ASIC physical tile.
When sufficient customer demand has accumulated for particular applications, multi project wafer (MPW) service [22] can be used to reduce the costs of low volume production of new ASIC physical tiles for implementing a virtual supercomputer for these applications. I.e., at each periodic run of the MPW service new popular customer applications collected and analyzed during the last time period can be included in the run.
The availability of
(iii) Recent advances in RTL to GDS-II conversion technologies for automated ASIC design;
Together open up new possibilities for creating energy efficient, high performance data centers based on ASIC physical tiles. Following this train of thought, an adaptive application-specific hardware lifetime management policy can be created, for allocating space to application-specific hardware in a data center, wherein the policy comprises the following steps:
The frequency of use of an (application, virtual supercomputer) pair can be measured, for example, as the ratio of the cumulative time spent in the virtual tiles of the said pair divided by the cumulative time spent in all applications in the last time period). The number of ASIC physical tiles installed in the data center should be proportional to the average frequency of use of the ASIC physical tile. But for important applications, the number of the physical tiles should be slightly higher than the average working set, in order to accommodate peak demand as well.
Of course, the data center cannot keep expanding with new hardware forever. Through time, the frequency of use of applications will change. To rebalance the allocation of data center space to different kinds of applications, less frequently used ASIC physical tiles can be periodically replaced by more frequently used ASIC physical tiles, according to the policy given above.
It is more practical to make the “field replacement unit” a rack module containing cloud building block chips, which in turn contain copies of a particular application-specific physical tile. Obsolete application-specific rack modules in the data center, which are no longer being used, will therefore be replaced over time, by application-specific rack modules for new customer applications.
Another way to distribute the physical tiles, which reduces the number of ASIC chips being released but increases the chip size, is to create a single chip kind, namely, a larger cloud building block chip that has, for example, a few physical tiles implementing A, some other physical tiles implementing B, some physical tiles realizing FPGA technology, and the remaining physical tiles implementing the “compiler-friendly general purpose supercomputer union chip” [Supercomputer]. In this case, the space on this large cloud building block chip can be allocated to applications using a similar application-specific hardware lifetime management policy over the generations of the chip. Each generation of the cloud building block will thus be tailored for applications currently considered important for the target customer community, as older generations of the cloud building block chip become obsolete.
3. Virtualizing Operating Systems
It suffices to make only a few changes to the baseline hypervisor system, in order to virtualize an entire operating system (OS) accelerated by a supercomputer, as opposed to just a user application accelerated by a supercomputer.
Here are some examples of the operation of the virtual hardware-accelerated OS: At the Ethernet connection of the OS placeholder tile, an inbound IP packet destined to the main IP address will be converted to a standard inter virtual-tile message from local virtual tile −1 to local virtual tile 0. The payload of a standard message sent from local virtual tile 0 to local virtual tile −1 will be sent out as an outbound IP packet by the OS placeholder tile, using the main IP address. A designated local virtual tile different from 0 can also communicate with the internet directly, by exchanging messages with local virtual tile −1. Local virtual tile −1 will forward inbound messages received using the secondary IP address of the Ethernet connection, to the designated local virtual tile different from 0. Also, an outbound message arriving from the designated local virtual tile different from 0, will be sent to the internet by local virtual tile −1, using the secondary IP address of the Ethernet connection.
At system initialization time, the saved initial state of local virtual tile 0 can represent an OS that has just been booted up, waiting for input from a remote main console, and the saved initial state of every other virtual tile can be idle, waiting for a message from the OS software to get started. When the microprocessor in local virtual tile 0 running the OS, arrives at an accelerated code fragment either in a user application or in kernel code, virtual tile 0 exchanges messages with other virtual tiles (e.g., virtual tile 1), thus initiating the actual hardware acceleration.
As an example of using the system, the performance critical parts a web service (such as a stock quote service) can be accelerated in this manner. The accelerated web service will appear as a user program within the virtual OS, where the user program has exclusive use of the secondary internet connection, and therefore all legacy software overheads of the OS for network accesses will be bypassed and replaced by parallel pipelined hardware serving to accelerate the complete web service as a whole. The frequent serializations due to user/kernel mode changes will be eliminated. Hardware resources of the virtual supercomputer implementing the web service can be incrementally increased or decreased over time at a virtual tile granularity, thus meeting cloud computing requirements.
This approach can boost performance through hardware acceleration of critical kernel and application code fragments, using a virtual supercomputer.
Some relevant difficulties of application-specific hardware acceleration of operating systems (e.g., precise exceptions including page faults, external and timer interrupts, privileged kernel code) were addressed in optimization 12 starting on p. 161, and optimization 13 starting on p. 166 of [Supercomputer]. I.e., it is possible to achieve hardware acceleration and yet retain binary compatibility with the original commodity OS software.
4. Alternative Physical Tile Replacement Algorithms for the Monitor Unit
The baseline version of the monitor unit runs a relatively simple physical tile replacement algorithm (the Least Recently Used algorithm). It is possible for the monitor unit to boost system performance, if it deploys a more advanced physical tile replacement algorithm.
As a more general replacement policy, each physical tile can be assigned a heuristic evaluation which is the weighted sum of a number of attributes of the physical tile, the virtual tile to be allocated to the physical tile, and several other system attributes. The physical tile which gets the highest heuristic evaluation is defined to be the best physical tile to replace.
An example of a monitoring algorithm is shown below. Upon a request for a replacement tile,
Several alternatives for speeding up the parallel implementation exist. For example:
The heuristic evaluations of each potential replacement tile can be based on the weighted sum of numerical measurements representative of the following features:
Reducing the communication latency among the virtual tiles of hardware accelerated applications: Every (application, virtual supercomputer) pair has a working set of one or more virtual tiles. In order to decrease the communication latency among the virtual tiles in a working set, the following rules should be applied. (i) Allocation of the first virtual tile: A set of physical tiles which are close together, with about the size of the working set (obtained by profiling earlier executions of the same application) will be reserved for this (application, virtual supercomputer) pair, if possible. The first virtual tile will preferably be allocated to a physical tile within the reserved set. (ii) Allocation of a virtual tile during normal operation: The virtual tile will preferably be assigned to a physical tile within the reserved set, which is close to the physical tiles presently belonging to the same virtual supercomputer.
Implementation of more advanced replacement policies: Based on the status update messages coming from the physical tiles, the monitor unit should continue to use the true LRU replacement policy when it works well. With dedicated hardware support the monitor unit can also use alternative replacement policies such as Least Frequently Used, and can switch to defensive replacement policies resilient to low reuse, when tile thrashing/low reuse is detected.
Re-using of physical tiles: It is possible to avoid the reconfiguration overhead of physical tiles. A virtual tile's state is composed of the configuration state (which specifies the function of the virtual tile) and the memory state (which is the current execution state including registers and SRAMs). Whenever a new virtual tile needs to be allocated, the Monitor unit should choose a physical tile that has already been configured with the configuration state of the new virtual tile.
Honoring service level agreements (SLAs): The monitor can differentiate the hardware accelerated applications based on their service level agreements. A physical tile that has been allocated to a virtual tile of an application with a “gold customer” SLA should have a less chance of being deallocated when it compares to the one that has been used by a virtual tile of an application with a “silver customer” or “bronze customer” SLA. More complex SLA rules, such as one involving monetary penalties for various levels of performance degradation can also be factored into the heuristic evaluation calculation, in an attempt to minimize losses to the data center operator.
Other optimizations, such as:
5. Avoiding Data Copying During Virtual Tile Migration
Notice that, following the non-virtual supercomputer design within [Supercomputer], the DRAM memory accesses done by a virtual tile (i.e., a design partition of the original non-virtual supercomputer) are always sent to the local DRAM unit near that virtual tile. A virtual supercomputer thus handles its own low-level DRAM data sharing among virtual tiles, e.g., memory coherence, through application communication messages sent between the virtual tiles. It is clear that the virtual tile state to be saved includes registers and SRAMs of the virtual tile, defining the current execution state of the virtual tile, and the configuration memory, defining the function of the virtual tile. The virtual tile execution state also includes the data structures in the local DRAM unit of the virtual tile. Local data structures can be read from an area within the local DRAM near the old physical tile where the virtual tile was, saved in hypervisor storage, and later restored from hypervisor storage to an area within the local DRAM near the new physical tile. Appropriate memory protection of one virtual supercomputer from another (such as a memory area range check) must be enforced. This approach will accurately reproduce the behavior of the original non-virtual supercomputer on a hypervisor system. When tile migrations occur infrequently, this state copying approach is simple and has high performance as well, since the physical tile will always remain near the DRAM unit it needs to access.
But, if state saving and restoring is too slow for data structures in DRAM, the DRAM resources in the hypervisor system can be consolidated as a single system-wide bank-interleaved shared memory. In this case, when a virtual tile v1 accesses DRAM, it will access the fixed memory area within the entire hypervisor system reserved for “the local DRAM of v1” (preferably in the DRAM unit near the first physical tile where v1 is allocated). When v1 is deallocated from a physical tile and later allocated to a different physical tile, the state of the “local DRAM of v1” memory altered by the first physical tile must be made available to the second physical tile where v1 is migrated, but the DRAM state need not be copied. The virtual tile will continue to access the same DRAM memory area from its new physical tile. In this case, reducing the distance between a physical tile and the DRAM units it needs to access, will be one of the heuristics used by the monitor.
6. Isolation Between Different (Application, Virtual Supercomputer) Pairs
In the main design of this document, we always treated a virtual tile as a pair (application id, local virtual tile number within this application's virtual supercomputer) so that messages from all virtual supercomputers could be routed in a uniform way. In order to enhance security, the application id part of the pair forming a virtual tile should not be written or read by the virtual supercomputer at all. The virtual supercomputer must communicate only with the virtual tiles of the same virtual supercomputer. This can be done by creating a wrapper module called an inner physical tile harness around the virtual tile within the physical tile, which cannot be accessed by the virtual tile except by pre-communication messages. The inner physical tile harness contains the application id register. Upon reset, the application id register of a normal physical tile is set to NULL. When a write_state request arrives at the physical tile, the application id register is also written from the “application id” part inside the state data being written. When an inbound pre-communication message arrives at the physical tile, the application id part of each global virtual tile field is verified to be equal to the application id register, and then removed to leave only the local virtual tile number. For outbound pre-communication messages, the application id is pre-pended to each of the local virtual tile number fields of the message coming out of the virtual tile.
Actually, to implement an inner physical tile harness, only a map from local virtual tile numbers within a virtual supercomputer to global virtual tile numbers encompassing all virtual tiles of all virtual supercomputers, and an inverse map for the same, is sufficient.
For example, alternatively, assuming the local virtual tiles of a virtual supercomputer are mapped to a contiguous area of the global virtual tile space, where the areas of different virtual supercomputers do not overlap, a unique virtual tile base register can be used in lieu of the application id register, where the virtual tile base is subtracted from the global virtual tile to obtain the corresponding local virtual tile when receiving an inbound pre-communication message, and where the virtual tile base is added to a local virtual tile to obtain a global virtual tile when sending an outbound pre-communication message. The virtual tile base register will be rewritten during each write_state request.
Please also see the next section, regarding how each user-level software application running on a host machine can be constrained by its OS, to exchange messages only with the local virtual tiles of its own virtual supercomputer.
7. Starting and Ending Virtual Supercomputer Execution
Notice that we did not mention how to insert a (software application, virtual supercomputer) pair into the hypervisor system, or how to remove a (software application, virtual supercomputer) pair from the hypervisor system. Thus, the system so far described is suitable for a continuously running cloud computing system with a fixed set of “approved” (application, virtual supercomputer) pairs.
Here, will describe a method for the creation of new (application, virtual supercomputer) pairs in the hypervisor system, and the destruction of such pairs.
For reducing the security risks in hardware designs (see, e.g., [23]) we recommend creating cryptographically signed initial states of virtual tiles that are generated using authorized tools, and registering the initial states of all virtual tiles of all virtual supercomputers before they are used. Registering a virtual tile means: checking the signature validity of the initial state of a virtual tile and moving that initial state to the hypervisor storage.
A distinguished application and its virtual supercomputer called the supervisor will be introduced here. The supervisor application is privileged: the supervisor application does not have virtual tiles in its virtual supercomputer other the virtual tile −1, but can exchange messages with any virtual tile of any virtual supercomputer. The inner tile harness protection is disabled for the supervisor. The registration of a new virtual supercomputer is done using a dedicated PCI Express connection to a secure host computer, or an encrypted Internet connection to a trusted remote server. Registration consists of inserting the clean initial state of each virtual tile v of each newly introduced virtual supercomputer in the hypervisor storage, by sending the following messages from tile −1 of the supervisor virtual supercomputer, over the outbound pre-communication network:
At system initialization time, virtual tile −1 of the supervisor supercomputer, pinned in a physical tile and serving as a message exchange gateway with the trusted server, will attempt to send a message to virtual tile v of the application. Since the destination virtual tile is initially not allocated in a physical tile, a local first level cache miss will occur in the supervisor virtual tile −1's physical tile harness. In this case, the physical tile harness of supervisor virtual tile −1 will recognize that (i) it is running the supervisor and that (ii) the “register” opcode is present in the message payload, and will forward the entire “register” message over the lookup network to the correct owner of virtual tile v, as follows:
Virtual tile v's owner unit will respond to the register request by:
Upon receiving the acknowledgement from the owner of v, the supervisor physical tile harness will have completed the registration operation. Then, an acknowledgement message is looped back to the physical tile containing the supervisor virtual −1 from its physical tile harness as follows:
The supervisor can consider the registration complete if and when it receives acknowledgements for each register request. As a result of registering, clean read-only copies of the initial state of virtual tiles will already exist in the hypervisor storage when any (application, virtual supercomputer) pair is started for the first time. The initial contents of a virtual tile implemented through a union chip ASIC physical tile, will be the configuration memory contents of the physical tile. If the physical tile is implemented with an FPGA, the initial state will be an FPGA configuration bitstream.
It makes sense to store only one copy of the initial state of virtual tiles of a given application, even though there may be multiple instances of the application running in the hypervisor system at a given time. For this purpose, it suffices to create a simple function to extract the application id for instance 0 of a given application, given the application id of any instance n of the same application. For example, the instance id may be the low order bits of the application id; therefore, the low order bits will be 0 for the case of instance 0. The application code should not have the privileges to read or write the application id field directly, it should exchange messages only with its locally numbered virtual tiles. To implement this constraint securely, the application id of the application is pre-pended to each message going from the application to the hypervisor system, automatically by a lightweight system call for message exchanges with the attached hypervisor system. In this manner, an instance of an application will be constrained to exchange messages only with the virtual tiles of its own virtual supercomputer and will not be able see or change its own application id.
The first time a virtual tile of a given (application, virtual supercomputer) is allocated in a physical tile, the writable state of the virtual tile will be missing. In this case, the allocation/deallocation FSM within the owner unit will create the initial writable state for this virtual tile of this instance of the application, by copying the configuration information from the clean read-only state of this virtual tile for instance 0 of the same application, and setting the writable part (registers, memories) of the virtual tile state to default initial values. Therefore, no special action is needed for initializing the virtual tiles when an (application, virtual supercomputer) pair starts.
However, as an (application, virtual supercomputer) pair ends, the hypervisor resources allocated to it (physical tiles that are still running, writable states of virtual tiles that were saved in hypervisor storage) must be released. This can be accomplished by issuing the following (user-level, non-privileged) message from the software application, for each virtual tile v of the virtual supercomputer:
From virtual tile −1 of the same application, just before the software application ends (e.g., these messages can be triggered in the software application by using an atexit call in a UNIX™-like system).
The physical tile harness of the application placeholder tile for the application understands the message contains a terminate request, and behaves as if a local first level cache miss occurred, for mapping virtual tile v to a physical tile, forwarding the terminate message to the owner of virtual tile v of the present application and instance, over the lookup network. The owner in turn forwards the terminate request to the allocation/deallocation FSM, which in turn checks if the virtual tile v is allocated in a physical tile p, and if so, issues a shutdown_and_read_state command to the physical tile p, but discards the state. Regardless of whether virtual tile v is allocated or not, the allocation/deallocation FSM also deletes the writable state for the virtual tile v from hypervisor storage, in case such a writable state record exists. As a result, all virtual tiles of this virtual supercomputer will be de-allocated, and all writable tile states of this virtual supercomputer will be deleted from hypervisor storage; thus achieving the termination of the virtual supercomputer.
The physical tile harness of local virtual tile −1 finally sends back an acknowledgement message corresponding to the terminate message back to the application, in order to assure that the software application can confirm completion of the virtual supercomputer activities before exiting from its process.
8. Heterogeneous Physical Tiles
The idea of application placeholder physical tiles can be easily generalized to N PCI Express connections supporting M>N applications. For example, when both a couple of instances of application A and an instance of application B are running on the same host processor and are communicating with their three respective virtual supercomputers with the same PCI Express connection, application placeholder virtual tiles −1 for the two instances of application A and also the application placeholder virtual tile −1 for application B may be implemented on the single physical tile attached to this single PCI Express connection. The system will behave as if three application placeholder sub-tiles have been implemented inside one single physical tile.
More generally, more than one virtual tile can be allocated inside sub-tiles within a single physical tile.
In a hypervisor system that includes sub-tiles, the following changes are required.
The owner data structures for mapping virtual tiles to physical tiles, and local caches within physical tile harnesses, will become mappings from virtual tiles to (physical tile, sub-tile) pairs. The monitor will supply (physical tile, sub-tile) pairs to preempt. The physical tile source and destination fields within messages will also be changed to pairs of the form (physical tile, sub-tile). However, routing from physical tile harnesses and to physical tile harnesses (e.g. within the communication, control and lookup networks) routing will still be done based on the physical tile portion of the (physical tile, sub-tile) pairs. Once an inbound message going to a (physical tile, sub-tile) enters the physical tile harness, and then reaches the inbound pre-communication channel, or the pre-control channel, the sub-tile part of the destination must be retained in the message for internal routing purposes within the physical tile, until the specified destination sub-tile within the physical tile is reached. Inner tile harnesses for hiding the application id register from the virtual tile are still needed for each sub-tile for security, but will now be called inner sub-tile harnesses.
Sub-tile addressing allows flexible allocation of virtual tiles to hardware resources if, for example, sub-tiles are composed of one or more contiguous hardware blocks of minimum size. For example, assuming a physical tile has 8 minimal sized blocks and sufficient reconfiguration capability, 8 sub-tiles of 1 block each (starting at blocks 0, 1, 2, 3, 4, 5, 6, 7), 4 sub-tiles of 2 blocks each (starting at blocks 0, 2, 4, 6), 2 sub-tiles of 4 blocks each (starting at blocks 0 and 4), or 1 sub-tile of 8 blocks (starting at 0), are some possibilities which can be implemented within this physical tile, using algorithms resembling the dynamic allocation of memory blocks.
Having heterogeneous physical sub-tiles in the platform requires that the monitor unit be modified to apply a matching filter to all physical sub-tiles before they are evaluated in terms of other possible criteria. That is, the matching filter shall mark a physical sub-tile as feasible if and only if it has the required resources to contain the virtual tile. Then, the monitor unit shall use only the feasible physical sub-tiles in the physical sub-tile replacement algorithm.
9. Increased Reliability
Hardware reliability is becoming an increasingly important issue, due to increased vulnerability to particle-induced soft errors and intermittent timing faults due to aging effects and voltage droop. Similarly, persistent timing faults caused by manufacturing variability and hard errors due to wear-out are becoming increasingly common. The proposed approach for virtualizing application-specific supercomputers provides numerous opportunities for improved fault detection and fault recovery.
The hypervisor system itself is a mainly manual hardware design consisting of components (mostly FSMs) and networks (such as butterfly and hypercube networks). The physical tile is not a simple FSM, it is in fact the most complex component of the hypervisor. Each physical tile in turn contains internal components and networks; but the physical tile will usually be generated by a compiler from sequential code [Supercomputer]. In both the compiler-generated and manual hardware designs, the techniques to achieve reliability are similar. We will review some of the techniques for achieving reliability here, with a sufficient level of detail, so that the integration of each of these reliability techniques in a compiler algorithm for generating hardware from sequential code, also becomes clear.
First, to achieve the detection of and recovery from soft errors, it is desirable to have a checkpoint-restart mechanism to be able to retry the hardware execution of a code fragment, when a potential soft error is detected in the code fragment. Here is a speculation/retry model for an operation (x,MEM)=ƒ(x,MEM) (where ƒ is either a simple operation, or a complex function call, or an inner loop nest in the region hierarchy of the program), which reads a memory MEM and a register x, and then writes the same memory MEM and register x. To be able to retry ƒ, we must first identify the memories and registers that are live at the retry point at the beginning of the invocation of ƒ (register x and memory MEM in this case), and revise ƒ to make it ƒ_speculative, to ensure the only the new versions of such memories and registers are written, so that the original memory and register inputs to ƒ are not clobbered when a soft error is detected and a retry occurs. When a soft error is detected (e.g. a mismatch is detected during a dual modular redundancy run of ƒ_speculative, or an unrecoverable ECC error occurs during ƒ_speculative) the ƒ_speculative invocation immediately returns with a condition code cc that is false, otherwise it returns with a condition code cc that is true, with the results in x′ and MEM′. If there is any error (cc is false), the speculative code fragment should be retried, if not, the results of the speculative code fragment should be committed, while still checking them for integrity/ECC errors.
The baseline hardware acceleration of a software application in [Supercomputer] already works like the speculation/retry model given above at a very coarse grain, where the function ƒ is the entire accelerated code fragment. The application-specific supercomputer has a large DRAM memory serving as a last level cache (the application address space is the root of the memory hierarchy). The modified lines of this last level cache are not written back to the software application memory until the end of the accelerated program fragment, at which point a “flush all dirty lines” request is issued by the accelerator. For an accelerator with dual modular redundancy and ECC in its last level cache, if a comparison mismatch or an unrecoverable ECC error is detected before reaching the point of flushing the dirty lines in the last level cache, it is possible to recover from the potential soft error by just discarding the accelerator state and restarting the entire accelerated code fragment from the beginning. The final commit operation (since it is not inside yet another checking harness) can be implemented with triple modular redundancy. The ECC of the results being committed to the application memory address space can be checked, and the data can be corrected if possible. If an unrecoverable ECC error occurs during the final committing of results, or if there are too many unsuccessful retries, the result will be a fatal error that should be reported from the virtual supercomputer to the software application, which should revert to software-only execution (the original software code will still be around). However, the offending physical tile and offending DRAM resources should be avoided in future runs.
In case a soft error is highly probable during a long-running accelerated code fragment, sub-regions smaller than the entire accelerated code fragment in the program region hierarchy can be speculatively executed by following the recipe in the speculation/retry model for soft errors given above.
The conventional approach for fault detection is to replicate hardware and compare the results (dual-modular redundancy). This approach can be realized by building redundancy into the FSM when creating the FSM. While a register to register operation is executed in duplicate, the two copies of each of the input operands should be verified to be equal. The FSM state transition logic should similarly be duplicated and at the beginning of each cycle/state the two copies of condition codes and state registers should be verified to be equal. ECC or parity should be generated and checked during memory operations. Checksums or other redundancy techniques can be used during network message transmissions.
A cheaper alternative technique is to use modulo N arithmetic (for a small N) for checking individual operations instead of full dual modular redundancy. If profiling data from a soft error simulation is available, checking logic can be implemented for the registers, functional units and FSM state transitions that are most prone to soft errors until an area budget reserved for reliability is exceeded.
Since the virtualized hardware is usually generated automatically from a high-level specification such as sequential code, further optimizations to reduce checking logic are also possible. Simplified versions of the hardware can be instantiated to check certain invariant properties of the physical tile's operation. These invariants can be explicitly provided by the programmer in the original code (programmer assertions often offer an independent check for results), or that can be inferred from the sequential code, for example by selecting a few among the assertions automatically generated by symbolic execution [Supercomputer]. For example, in the above speculation/retry model example, the ƒ computation can be a sorting algorithm (without dual modular redundancy), and the verification computation can be a check that a random subsequence of the array is ordered. If this simple check fails, the sorting routine is retried; but if it succeeds, the state changes produced by the sorting routine are committed.
The probability of failure throughout the system can also be minimized by conventional circuit-level hardening (for soft errors), and wear-leveling (for aging-induced transient and permanent failures).
An end-to-end checksum is often a more hardware-efficient technique for networks. When a message with a wrong checksum arrives into any FSM, a speculation failure action may be performed.
Permanent failures in the network can also be detected, and can be rectified by disabling failed nodes and reconfiguring the packet routing logic to avoid such nodes. This is only possible with network topologies that provide path redundancy (i.e., more than one possible route from each source to each destination).
The invention has been shown and described with reference to a particular preferred embodiment. However, it is to be understood that the invention is not limited to that particular embodiment, and that various modifications, additions and alterations may be made to the invention by one skilled in the art without departing from the spirit and scope of the invention.
This application claims priority, as a continuation application, to U.S. patent application Ser. No. 16/105,741 filed on Aug. 20, 2018, which claims priority, as a continuation application to U.S. patent application Ser. No. 15/137,268 filed on Apr. 25, 2016, now U.S. Pat. No. 10,120,704, which claims priority, as a continuation application, to U.S. patent application Ser. No. 13/366,318 filed on Feb. 4, 2012, now U.S. Pat. No. 9,465,632. Documents Ser. Nos. 16/105,741, 15/137,268, 13/366,318 and U.S. Pat. Nos. 10,120,704, 9,465,632 are hereby incorporated by reference.
Number | Name | Date | Kind |
---|---|---|---|
6034538 | Abramovici | Mar 2000 | A |
6826615 | Barrall et al. | Nov 2004 | B2 |
7409670 | Pritchard et al. | Aug 2008 | B1 |
7555566 | Blumrich et al. | Jun 2009 | B2 |
7603540 | Doering et al. | Oct 2009 | B2 |
7734894 | Wentzlaff | Jun 2010 | B1 |
7761687 | Blumrich et al. | Jul 2010 | B2 |
7814486 | Papakipos et al. | Oct 2010 | B2 |
8020163 | Nollet et al. | Sep 2011 | B2 |
20040049672 | Nollet et al. | Mar 2004 | A1 |
20090187756 | Nollet | Jul 2009 | A1 |
20090287571 | Fujioka | Nov 2009 | A1 |
20100153966 | Arimilli et al. | Jun 2010 | A1 |
20110072180 | Lee | Mar 2011 | A1 |
20110173622 | Shin | Jul 2011 | A1 |
20110307663 | Kultursay et al. | Dec 2011 | A1 |
20130005068 | Nickel et al. | Jan 2013 | A1 |
Number | Date | Country |
---|---|---|
1574965 | Sep 2005 | EP |
Entry |
---|
Chun-Hsian Huang, “Software-Controlled Dynamically Swappable Hardware Design in Partially Reconfigurable Systems” (Year: 2007). |
Ming Liu, Wolfgang Kuehn, Zhonghai Lu, and Axel Janisch, “Run-time Partial Reconfiguration Speed Investigation and Architectural Design Space Exploration”, In Proceedings of the International Conference on Field Programmable Logic and Applications 2009 (FPL'09), Prague, Czech Republic, Aug. 2009. |
Fabrizio Ferrand!, Marco D. Santambrogio, Donatella Sciuto, “A Design Methodology for Dynamic Reconfiguration: The Caronte Architecture,” ipdps, vol. 4, pp. 163b, 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS'05)—Workshop 3, 2005. |
Mateusz Majer, Juergen Teich, Ali Ahmadinia, and Christophe Bobda. 2007. The Erlangen Slot Machine: A Dynamically Reconfigurable FPGA-based Computer. J. VLSI Signal Process. Sys!. 47, 1 (Apr. 2007), 15-31. DOI=10.1007/s11265-006-0017-6 http://dx.doi.org/10.1007/s11265-006-0017-6. |
Chun-Hsian Huang and Pao-Ann Hsiung. “Software-Controlled Dynamically Swappable Hardware Design in Partially Reconfigurable Systems,” EURASIP Journal on Embedded Systems, vol. 2008, pp. 1-11, 2008, doi:10.1155/2008/231940. |
Vincenzo Rana, Marco Santambrogio, Donatella Sciuto, Boris Kettelhoit, Markus Koester, Mario Porrmann, Ulrich Rucker!, “Partial Dynamic Reconfiguration in a Multi-FPGA Clustered Architecture Based on Linux,” ipdps, pp. 173, 2007 IEEE International Parallel and Distributed Processing Symposium, 2007. |
Klaus Danne and Marco Platzner. Periodic real-time scheduling for FPGA computers. Third IEEE International Workshop on Intelligent Solutions in Embedded Systems (WISES'05), Hamburg University of Technology, May 2005. http://dx.doi.org/10.1109/WISES.2005.1438720. |
H. Simmler and L. Levinson and R. Manner. Multitasking on FPGA Coprocessors. Proceedings of the 1oth International Workshop on Field Programmable Gate Arrays (FPL). pp. 121-130. 2000. Springer. |
Miljan Vuletic, Laura Pozzi, and Paolo Ienne. 2005. Seamless Hardware-Software Integration in Reconfigurable Computing Systems. IEEE Design and Test of Computers, vol. 22, No. 2 (Mar. 2005), pp. 102-113. DOI=10.1109/MDT.2005.44 http://dx.doi.org/10.1109/MDT.2005.44. |
Hayden Kwok-Hay So, Artem Tkachenko, Robert Brodersen. “A Unified Hardware/Software Runtime Environment for FPGA-Based Reconfigurable Computers using BORPH”. Proceedings of the 4th International Conference on Hardware/Software Codesign and System Synthesis, 259-264, 2006. |
Eray Ozkural, A two-dimensional ISA, service invention report, Erendiz Superbilgisayar Ltd., Eskisehir, Turkey, Jan. 6, 2009. |
Co-owned, co-pending U.S. Appl. No. 13/296,232, filed Nov. 15, 2011, Ebcioglu et al., Method and system for converting a single-threaded software program into an application-specific supercomputer. |
Withdrawal of EPO appeal dated Jul. 16, 2020, Appeal No. T0923/19-3.5.06, EPO application No. EP12867516, claiming priority of U.S. Appl. No. 13/366,318 (1 page). |
Applicant's arguments and amendments before the EPO dated Jul. 20, 2020, EPO application No. EP12867516, claiming priority of U.S. Appl. No. 13/366,318 (31 pages). |
Shimada, Toshio, Toda, Kenji, Nishida, Kenji. “Real-Time Parallel Architecture for Sensor Fusion” Journal of Parallel and Distributed Computing 15, 143-152 (1992) http://dx.doi.org/10.1016/0743-7315(92)90106-W. |
European Patent Office, Extended European Search Report dated Apr. 29, 2016 for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318. |
Marescaux T et al: “Networks on Chip as Hardware Components of an OS for Reconfigurable Systems” International Conference on Field-Programmable Logic, FPL 2003, Lecture Notes in Computer Science, vol. 2778, Sep. 1, 2003 (Sep. 1, 2003 ), pp. 595-605, XP002319855, ISBN: 978-3-540-45234-8. |
Rosti E et al: “Robust partitioning policies of multiprocessor systems”, Performance Evaluation, Amsterdam, NL, vol. 19, No. 2-3, Mar. 1994 (Mar. 1994), pp. 141-165, XP026655278, ISSN: 0166-5316, DOI: 10.1016/0166-5316 (94)90037-X. |
Japan Patent Office, “Decision to Grant a Patent” dated May 24, 2016 for application JP 2014-555553 PCT/US2012072200 claiming priority of U.S. Appl. No. 13/366,318. |
EPO Communication pursuant to Article 94(3) EPC dated Dec. 12, 2018, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (4 pages). |
Applicant's statement of grounds of appeal dated Feb. 15, 2019, against EPO application EP 12867516 interlocutory decision dated Nov. 14, 2018 (28 pages). |
Applicant's response dated Feb. 20, 2019 to EPO Communication pursuant to Article 94(3) EPC dated Dec. 12, 2018, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (13 pages). |
Nishida, Kenji, Toda, Kenji, Shimada, Toshio, Yamaguchi, Yoshinori. “The Hardware Architecture of the CODA Real-Time Processor”, Proceedings of PARCO 93, Parallel Computing: Trends and Applications, G.R. Joubert, D. Trystram, F.J. Peters, D.J. Evans (editors), 1994 Elsevier Science B.V. |
Applicant's response dated Oct. 12, 2016 to EPO Extended European Search Report dated Apr. 29, 2016 for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (53 pages). |
EPO communication pursuant to Article 94(3) EPC dated Feb. 6, 2017, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (5 pages). |
Applicant's response dated May 10, 2017 to EPO communication pursuant to Article 94(3) EPC dated Feb. 6, 2017, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (67 pages). |
Applicant's response dated Feb. 5, 2020 to EPO summons to oral proceedings pursuant to Rule 115(1) EPC dated Nov. 28, 2019, Appeal No. T0923/19-3.5.06, EPO application No. EP12867516, claiming priority of U.S. Appl. No. 13/366,318 (49 pages). |
EPO application EP 12867516.2 interlocutory decision refusing the refund of further search fees paid according to Rule 164(1) EPC, dated Nov. 14, 2018. |
EPO summons to oral proceedings pursuant to Rule 115(1) EPC dated Nov. 28, 2019, Appeal No. T0923/19-3.5.06, application No. EP12867516, claiming priority of U.S. Appl. No. 13/366,318 (18 pages). |
Nishida, Kenji, Toda, Kenji, Takahashi, Eiichi, Yamaguchi, Yoshinori. “An Architecture of the Real-Time Parallel Processor CODA,” Journal of Institute of Electronics, Information and Communication Engineers, vol. J78-D-I No. 8, Aug. 1995, pp. 777-787. |
Marescaux, T., Nollet, V., Mignolet, J.Y., Bartic, A., Moffat, W., Avasarc, P., Coene, P., Verkest, D., Vernalde, S., Lauwereins, R . . . “Run-time support for heterogenous multitasking on reconfigurable SoCs” Integration, The VLSI Journal, North-Holland Publishing Company. Amsterdam, NL, vol. 38, No. 1, Oct. 1, 2004, pp. 107-130. North Holland. XP004641567. ISSN: 9167-9260. DOI: 10.1016/J.VLSI.2004.03.002. |
Japan Patent Office, First Office Action for application JP 2014-555553, dated Nov. 26, 2015. From PCT version of present application PCT/US2012/072200. |
European Patent Office, Supplementary Partial European Search Report for application EP 12867516.2, dated Nov. 27, 2015. From PCT version of present application PCT/US2012/072200. |
EPO Communication pursuant to Article 94(3) EPC dated Dec. 1, 2017, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (6 pages). |
Applicant's response dated Mar. 22, 2018 to the EPO Communication pursuant to Article 94(3) EPC dated Dec. 1, 2017, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (24 pages). |
Wikipedia,“Reconfigurable Computing” article, Sep. 24, 2019. https://en.wikipedia.org/w/index.php?title=Reconfigurable_computing&oldid=917632607 (10 pages). |
Nikolaos Voros, Alberto Rosti, Michael Hubner (editors): Dynamic System Reconfiguration in Heterogeneous Platforms—The Morpheus Approach; Springer Verlag, Lecture Notes in Electrical Engineering 40, 2009. DOI: 10.1007/978-90-481-2427-5 (277 pages). |
Tarek El-Ghazawi, Esam El-Araby, Miaoqing Huang, Kris Gaj, Volodymyr Kindratenko, Duncan Buell, “The Promise of High-Performance Reconfigurable Computing”, Computer ( vol. 41 , Issue: 2 , Feb. 2008), pp. 69-76, https://doi.org/10.1109/MC.2008.65 (8 pages). |
EPO communication pursuant to Rule 137(4) EPC and Article 94(3) EPC dated Jul. 24, 2017, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (1 page). |
Applicant's response dated Aug. 25, 2017 to EPO communication pursuant to Rule 137(4) EPC and Article 94(3) EPC dated Jul. 24, 2017, for application EP12867516 PCT/US2012072200, claiming priority of U.S. Appl. No. 13/366,318 (21 pages). |
Minutes of EPO oral proceedings held on Jul. 15, 2020, issued on Jul. 23, 2020, Appeal No. T0923/19-3.5.06, EPO application No. EP12867516, claiming priority of U.S. Appl. No. 13/366,318 (4 pages). |
Number | Date | Country | |
---|---|---|---|
20190179656 A1 | Jun 2019 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 16105741 | Aug 2018 | US |
Child | 16280125 | US | |
Parent | 15137268 | Apr 2016 | US |
Child | 16105741 | US | |
Parent | 13366318 | Feb 2012 | US |
Child | 15137268 | US |