US20130318276A1

US20130318276A1 - Offloading of computation for rack level servers and corresponding methods and systems

Info

Publication number: US20130318276A1
Application number: US13/900,241
Authority: US
Inventors: Parin Bhadrik Dalal
Original assignee: Xockets IP LLC
Current assignee: XOCKETS Inc; Xockets IP LLC
Priority date: 2012-05-22
Filing date: 2013-05-22
Publication date: 2013-11-28
Also published as: US11080209B2; WO2013177316A2; US20130347110A1; WO2013177313A2; US9665503B2; US20170237624A1; WO2013177313A3; US9258276B2; US20130318275A1; US20130318119A1; WO2013177310A2; US20140157396A1; US9619406B2; US20130346469A1; US10223297B2; WO2013177316A3; WO2013177310A3; US9495308B2; US20170237714A1; US20130318084A1

Abstract

A system is disclosed that can include at least one processor module connectable to a memory bus. The processor module can include at least one memory, at least one offload processor mounted on the processor module, and configured to execute operations on data received over the memory bus, and to output context data to the memory and read context data from the memory, and a hardware scheduling logic mounted on the module and configured to control operations of the at least one processor.

Description

PRIORITY CLAIMS

This application claims the benefit of U.S. Provisional Patent Application 61/650,373 filed May 22, 2012, the contents of which are incorporated by reference herein.

TECHNICAL FIELD

The present disclosure relates generally to servers, and more particularly to offload or auxiliary processing modules that can be physically connected to a system memory bus to process data independent of a host processor of the server.

BACKGROUND

Networked applications often run on dedicated servers that support an associated “state” for context or session-defined application. Servers can run multiple applications, each associated with a specific state running on the server. Common server applications include an Apache web server, a MySQL database application, PHP hypertext preprocessing, video or audio processing with Kaltura supported software, packet filters, application cache, management and application switches, accounting, analytics, and logging.
Unfortunately, servers can be limited by computational and memory storage costs associated with switching between applications. When multiple applications are constantly required to be available, the overhead associated with storing the session state of each application can result in poor performance due to constant switching between applications. Dividing applications between multiple processor cores can help alleviate the application switching problem, but does not eliminate it, since even advanced processors often only have eight to sixteen cores, while hundreds of application or session states may be required.

SUMMARY

A system can include at least one processor module connectable to a memory bus, the processor module including; at least one memory, at least one offload processor mounted on the processor module, and configured to execute operations on data received over the memory bus, and to output context data to the memory and read context data from the memory, and a hardware scheduling logic mounted on the module and configured to control operations of the at least one processor

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows illustrates an embodiment with a group of web servers that are partitioned across a group of brawny processor core(s) and a set of wimpy cores housed in a rack server.

FIG. 2 shows an embodiment with an assembly that is favorably suited for handling real time traffic such as video streaming.

FIG. 3 shows illustrates an embodiment with a proxy server—web server assembly that is partitioned across a group of brawny processor core(s) (housed in a traditional server module) and a set of wimpy cores housed in a rack server module.

FIG. 4-1 shows a cartoon schematically illustrating a data processing system according to an embodiment, including a removable computation module for offload of data processing.

FIG. 4-2 shows an example layout of an in-line module (referred to as a “XIMM”) module according to an embodiment.

FIG. 4-3 shows two possible architectures for a data processing system including X86 main processors and XIMMs (Xockets MAX and MIN).

FIG. 4-4 shows a representative the power budget for XIMMs according to various embodiments.

FIG. 4-5 illustrates data flow operation of one embodiment of a XIMM using an ARM A9 architecture.

DETAILED DESCRIPTION

Networked applications are available that run on servers and have associated with them a state (session-defined applications). The session nature of such applications allows them to have an associated state and a context when the session is running on the server. Further, if such session-limited applications are computationally lightweight, they can be run in part or fully on the auxiliary or additional processor cores (such as those based on the ARM architecture, as but one particular example) which are mounted on modules connected to a memory bus, for example, by insertion into a socket for a Dual In-line Memory Module (DIMM). Such modules can be referred to as a Xocket ™In-line Memory Module (XIMM), and have multiple cores (e.g., ARM cores) associated with a memory channel. A XIMM can access the network data through an intermediary virtual switch (such as OpenFlow or similar) that can identify sessions and direct the network data to the corresponding module (XIMM) mounted cores, where the session flow for the incoming network data can be handled.
As will be appreciated, through usage of a large prefetch buffer or low latency memory, the session context of each of the sessions that are run on the processor cores of a XIMM can be stored external to the cache of such processor cores. By systematically engineering the transfer of cache context to a memory external to the module processors (e.g., RAMs) and engineering low latency context switch, it is possible to execute several high-bandwidth server applications on a XIMM provided the applications are not computationally intensive. The “wimpy” processor cores of a XIMM can be favorably disposed to handle high network bandwidth traffic at a lower latency and at a very low power when compared to traditional high power ‘brawny’ cores.
In effect, one can reduce problems associated with session limited servers by using the module processor (e.g., an ARM architecture processor) of a XIMM to offload part of the functionality of traditional servers. Module processor cores may be suited to carry computationally simple or lightweight applications such as packet filtering or packet logging functions. They may also be suited for providing the function of an application cache for handling hot-code that is to be serviced very frequently to incoming streams. Module processor cores can also be suited for functions such as video streaming/ real time streaming, that often only require light-weight processing.
As an example of partitioning applications between a XIMM with “wimpy” ARM cores and a conventional “brawny” core (e.g., x86 or Itanium server processor with Intel multicore processor), a computationally lightweight Apache web server can be hosted on one or more XIMMs with ARM cores, while computationally heavy MySQL and PHP are hosted on x86 brawny cores. Similarly, lightweight applications such as a packet filter, application cache, management and application switch are hosted on XIMM(s), while x86 cores host control, accounting, analytics and logging.
FIG. 1 illustrates an embodiment with a group of distributed web servers that are partitioned across a group of brawny processor core(s) 108 connected by bus 106 to switch 104 (which may be an OpenFlow or other virtual switch) and a set of wimpy XIMM mounted cores (112 a to 112 c), all being housed in a rack server module 140. In some embodiments, a rack server module 140 further includes a switch (100), which can be a network interface card with single root IO virtualization that provides input-out memory management unit (IOMMU) functions 102. A second virtual switch (104) running, for example, an open source software stack including OpenFlow can redirect packets to XIMM mounted cores (112 a to 112 c).
According to some embodiments, a web server running Apache-MySQL-PHP (AMP) can be used to service clients that send requests to the server module 140 from network 120. The embodiment of FIG. 1 can split a traditional server module running AMP across a combination of processors cores, which act as separate processing entities. Each of the wimpy processor cores (112 a to 112 c) (which can be low power ARM cores in particular embodiments) can be mounted on an XIMM, with each core being allocated a memory channel (110 a, 110 b, 110 c). At least of one of the wimpy processor cores (112 a to 112 c) can be capable of running a computationally light weight Apache or similar web server code for servicing client requests which are in the form of HTTP or a similar application level protocol. The Apache server code can be replicated for a plurality of clients to service a huge number of requests. The wimpy cores (112 a to 112 c) can be ideally suited for running such Apache code and responding to multiple client requests at a low latency. For static data that is available locally, wimpy cores (112 a to 112 c) can lookup such data from their local cache or a low latency memory associated with them. In case the queried data is not available locally, the wimpy cores (112 a to 112 c) can request a direct memory access (DMA) (memory-to-memory or disk-to-memory) transfer to acquire such data.
The computation and dynamic behavior associated with the web pages can be rendered by PHP or such other server side scripts running on the brawny cores 108. The brawny cores might also have code/scripting libraries for interacting with MySQL databases stored in hard disks present in said server module 140. The wimpy cores (112 a to 112 c), on receiving queries or user requests from clients, transfer embedded PHP/MySQL queries to said brawny cores over a connection (e.g., an Ethernet-type connection) that is tunneled on a memory bus such as a DDR bus. The PHP interpreter on brawny cores 108 interfaces and queries a MySQL database and processes the queries before transferring the results to the wimpy cores (112 a to 112 c) over said connection. The wimpy cores (112 a to 112 c) can then service the results obtained to the end user or client.
Given that the server code lacking server side script is computationally light weight, and many Web API types are Representational State Transfer (REST) based and require only HTML processing, and on most occasions require no persistent state, wimpy cores (112 a-112 c) can be highly suited to execute such light weight functions. When scripts and computation is required, the computation is handled favorably by brawny cores 108 before the results are serviced to end users. The ability to service low computation user queries with a low latency, and the ability to introduce dynamicity into the web page by supporting server-side scripting make the combination of wimpy and brawny cores an ideal fit for traditional web server functions. In the enterprise and private datacenter, simple object access protocol (SOAP) is often used, making the ability to context switch with sessions performance critical, and the ability of wimpy cores to save the context in an extended cache can enhance performance significantly.
FIG. 2 illustrates an embodiment with an assembly that is favorably suited for handling real time traffic such as video streaming. The assembly comprises of a group of web servers that are partitioned across a group of brawny processor core(s) 208 and a set of wimpy cores (212 a to 212 c) housed in a rack server module 240. The embodiment of FIG. 2 splits a traditional server module capable of handling real time traffic across a combination of processors cores, which act as separate processing entities. In some embodiments, a rack server module 240 further includes a switch (100), which can provide input-out memory management unit (IOMMU) functions 102.
Each of the wimpy processor cores (e.g., ARM cores) (212 a to 212 c) can be mounted on an in-memory module (not shown) and each of them can be allocated a memory channel (210 a to 210 c). At least one of the wimpy processor cores (212 a to 212 c) can be capable of running a tight, computationally light weight web server code for servicing applications that need to be transmitted with a very low latency/jitter. Example applications such as video, audio, or voice over IP (VoIP) streaming involve client requests that need to be handled with as little latency as possible. One particular protocol suitable for the disclosed embodiment is Real-Time Transport Protocol (RTP), an Internet protocol for transmitting real-time data such as audio and video. RTP itself does not guarantee real-time delivery of data, but it does provide mechanisms for the sending and receiving applications to support streaming data.
Brawny processor core(s) 208 can be connected by bus 206 to switch 204 (which may be an OpenFlow or other virtual switch). In one embodiment, such a bus 206 can be a front side bus.
In operation, server module 240 can handle several client requests and services information in real time. The stateful nature of applications such as RTP/video streaming makes the embodiment amenable to handle several queries at a very high throughput. The embodiment can have an engineered low latency context overhead system that enables wimpy cores (212 a to 212 c) to shift from servicing one session to another session in real time. Such a context switch system can enable it to meet the quality of service (QoS) and jitter requirements of RTP and video traffic. This can provide substantial performance improvement if the overlay control plane and data plane (for handling real time applications related traffic) is split across a brawny processor 208 and a number of wimpy cores (212 a to 212 c). The wimpy cores (212 a to 212 c) can be favorably suited to handling the data plane and servicing the actual streaming of data in video/audio streaming or RTP applications. The ability of wimpy cores (212 a to 212 c) to switch between multiple sessions with low latency makes them suitable for handling of the data plane.
For example, wimpy cores (212 a to 212 c) can run code that quickly constructs data that is in an RTP format by concatenating data (that is available locally or through direct memory access (DMA) from main memory or a hard disk) with sequence number, synchronization data, timestamp etc., and sends it over to clients according to a predetermined protocol. The wimpy cores (212 a to 212 c) can be capable of switching to a new session/new client with a very low latency and performing a RTP data transport for the new session. The brawny cores 208 can be favorably suited for overlay control plane functionality.
The overlay control plane can often involve computationally expensive actions such as setting up a session, monitoring session statistics, and providing information on QoS and feedback to session participants. The overlay control plane and the data plane can communicate over a connection (e.g., an Ethernet-type connection) that is tunneled on a memory bus such as a DDR bus. Typically, overlay control can establish sessions for features such as audio/videoconferencing, interactive gaming, and call forwarding to be deployed over IP networks, including traditional telephony features such as personal mobility, time-of-day routing and call forwarding based on the geographical location of the person being called. For example, the overlay control plane can be responsible for executing RTP control protocol (RTCP, which forms part of the RTP protocol used to carry VoIP communications and monitors QoS); Session Initiation Protocol (SIP, which is an application-layer control signaling protocol for Internet Telephony); Session Description Protocol (SDP, which is a protocol that defines a text-based format for describing streaming media sessions and multicast transmissions); or other low latency data streaming protocols.
FIG. 3 illustrates an embodiment with a proxy server—web server assembly that is partitioned across a group of brawny processor core(s) 328 (housed in a traditional server module 360) and a set of wimpy cores (312 a to 312 c) housed in a rack server module 340. The embodiment can include a proxy server module 340 that can handle content that is frequently accessed. A switch/load balancer apparatus 320 can direct all incoming queries to the proxy server module 340. The proxy server module 340 can look up its local memory for frequently accessed data and responds to the query with a response if such data is available. The proxy server module 340 can also store server side code that is frequently accessed and can act as a processing resource for executing the hot code. For queries that are not part of the rack hot code, the wimpy cores (312 a to 312 c) can redirect the traffic to brawny cores (308, 328) for processing and response.
In particular embodiments, in some embodiments, a rack server module 240 further includes a switch (100), which can provide input-out memory management unit (IOMMU) functions 302 and a switch 304 (which may be an OpenFlow or other virtual switch). Brawny processor core(s) 308 can be connected to switch 304 by bus 306, which can be a front side bus. A traditional server module 360 can also include a switch 324 can provide IOMMU functions 326.
The following example(s) provide illustration and discussion of exemplary hardware and data processing systems suitable for implementation and operation of the foregoing discussed systems and methods. In particular hardware and operation of wimpy cores or computational elements connected to a memory bus and mounted in DIMM or other conventional memory socket is discussed.
FIG. 4-1 is a cartoon schematically illustrating a data processing system 400 including a removable computation module 402 for offload of data processing from x86 or similar main/server processors 403 to modules connected to a memory bus 403.
Such modules 402 can be XIMM modules, as described herein or equivalents, and can have multiple computation elements that can be referred to as “offload processors” because they offload various “light touch” processing tasks such HTML, video, packet level services, security, or data analytics. This is of particular advantage for applications that require frequent random access or application context switching, since many server processors incur significant power usage or have data throughput limitations that can be greatly reduced by transfer of the computation to lower power and more memory efficient offload processors.
The computation elements or offload processors can be accessible through memory bus 405. In this embodiment, the module can be inserted into a Dual Inline Memory Module (DIMM) slot on a commodity computer or server using a DIMM connector (407), providing a significant increase in effective computing power to system 400. The module (e.g., XIMM) may communicate with other components in the commodity computer or server via one of a variety of busses including but not limited to any version of existing double data rate standards (e.g., DDR, DDR2, DDR3, etc.)
This illustrated embodiment of the module 402 contains five offload processors (400 a, 400 b, 400 c, 400 d, 400 e) however other embodiments containing greater or fewer numbers of processors are contemplated. The offload processors (400 a to 400 e) can be custom manufactured or one of a variety of commodity processors including but not limited to field-programmable grid arrays (FPGA), microprocessors, reduced instruction set computers (RISC), microcontrollers or ARM processors. The computation elements or offload processors can include combinations of computational FPGAs such as those based on Altera, Xilinx (e.g., Artix™ class or Zynq® architecture, e.g., Zynq® 7020), and/or conventional processors such as those based on Intel Atom or ARM architecture (e.g., ARM A9). For many applications, ARM processors having advanced memory handling features such as a snoop control unit (SCU) are preferred, since this can allow coherent read and write of memory. Other preferred advanced memory features can include processors that support an accelerator coherency port (ACP) that can allow for coherent supplementation of the cache through an FPGA fabric or computational element.
Each offload processor (400 a to 400 e) on the module 402 may run one of a variety of operating systems including but not limited to Apache or Linux. In addition, the offload processors (400 a to 400 e) may have access to a plurality of dedicated or shared storage methods. In this embodiment, each offload processor can connect to one or more storage units (in this embodiments, pairs of storage units 404 a, 404 b, 404 c and 404 d). Storage units (404 a to 404 d) can be of a variety of storage types, including but not limited to random access memory (RAM), dynamic random access memory (DRAM), sequential access memory (SAM), static random access memory (SRAM), synchronous dynamic random access memory (SDRAM), reduced latency dynamic random access memory (RLDRAM), flash memory, or other emerging memory standards such as those based on DDR4 or hybrid memory cubes (HMC).
FIG. 4-2 shows an example layout of a module (e.g., XIMM) such as that described in FIG. 4-1, as well as a connectivity diagram between the components of the module. In this example, five Xilinx™ Zynq® 7020 (416 a, 416 b, 416 c, 416 d, 416 e and 416 in the connectivity diagram) programmable systems-on-a-chip (SoC) are used as computational FPGAs/offload processors. These offload processors can communicate with each other using memory-mapped input-output (MMIO) (412). The types of storage units used in this example are SDRAM (SD, one shown as 408) and RLDRAM (RLD, three shown as 406 a, 406 b, 406 c) and an Inphi™ iMB02 memory buffer 418. Down conversion of 3.3 V to 2.5 volt is required to connect the RLDRAM (406 a to 406 c) with the Zynq® components. The components are connected to the offload processors and to each other via a DDR3 (414) memory bus. Advantageously, the indicated layout maximizes memory resources availability without requiring a violation of the number of pins available under the DIMM standard.
In this embodiment, one of the Zynq® computational FPGAs (416 a to 416 e) can act as arbiter providing a memory cache, giving an ability to have peer to peer sharing of data (via memcached or OMQ memory formalisms) between the other Zynq® computational FPGAs (416 a to 416 e). Traffic departing for the computational FPGAs can be controlled through memory mapped I/O. The arbiter queues session data for use, and when a computational FPGA asks for address outside of the provided session, the arbiter can be the first level of retrieval, external processing determination, and predictors set.
FIG. 4-3 shows two possible architectures for a module (e.g., XIMM) in a simulation (Xockets MAX and MIN). Xockets MIN (420 a) can be used in low-end public cloud servers, containing twenty ARM cores (420 b) spread across fourteen DIMM slots in a commodity server which has two Opteron x86 processors and two network interface cards (NICs) (420 c). This architecture can provide a minimal benefit per Watt of power used. Xockets MAX (422 a) contains eighty ARM cores (422 b) across eight DIMM slots, in a server with two Opteron x86 processors and four NICs (422 c). This architecture can provide a maximum benefit per Watt of power used.
FIG. 4-4 shows a representative power budget for an example of a module (e.g., XIMM) according to a particular embodiment. Each component is listed (424 a, 424 b, 424 c, 424 d) along with its power profile. Average total and total wattages are also listed (426 a, 426 b). In total, especially for I/O packet processing with packet sizes on the order 1 KB in size, module can have a low average power budget that is easily able to be provided by the 22 V_ddpins per DIMM. Additionally, the expected thermal output can be handled by inexpensive conductive heat spreaders, without requiring additional convective, conductive, or thermoelectric cooling. In certain situations, digital thermometers can be implemented to dynamically reduce performance (and consequent heat generation) if needed.
Operation of one embodiment of a module 430 (e.g., XIMM) using an ARM A9 architecture is illustrated with respect to FIG. 4-5. Use of ARM A9 architecture in conjunction with an FPGA fabric and memory, in this case shown as reduced latency DRAM (RLDRAM) 438, can simplify or makes possible zero-overhead context switching, memory compression and CPI, in part by allowing hardware context switching synchronized with network queuing. In this way, there can be a one-to-one mapping between thread and queues. As illustrated, the ARM A9 architecture includes a Snoop Control Unit 432 (SCU). This unit allows one to read out and write in memory coherently. Additionally, the Accelerator Coherency Port 434 (ACP) allows for coherent supplementation of the cache throughout the FPGA 436. The RLDRAM 438 provides the auxiliary bandwidth to read and write the ping-pong cache supplement (435): Block1$ and Block2$ during packet-level meta-data processing.
The following table (Table 1) illustrates potential states that can exist in the scheduling of queues/threads to XIMM processors and memory such as illustrated in FIG. 4-5.

TABLE 1

Queue/Thread State	HW treatment

Waiting for Ingress	All ingress data has been processed and thread
Packet	awaits further communication.
Waiting for MMIO	A functional call to MM hardware (such as
	HW encryption or transcoding) was made.
Waiting for Rate-limit	The thread's resource consumption exceeds limit,
	due to other connections idling.
Currently being	One of the ARM cores is already processing this
processed	thread, cannot schedule again.
Ready for Selection	The thread is ready for context selection.

These states can help coordinate the complex synchronization between processes, network traffic, and memory-mapped hardware. When a queue is selected by a traffic manager a pipeline coordinates swapping in the desired L2 cache (440), transferring the reassembled IO data into the memory space of the executing process. In certain cases, no packets are pending in the queue, but computation is still pending to service previous packets. Once this process makes a memory reference outside of the data swapped, a scheduler can require queued data from a network interface card (NIC) to continue scheduling the thread. To provide fair queuing to a process not having data, the maximum context size is assumed as data processed. In this way, a queue must be provisioned as the greater of computational resource and network bandwidth resource, for example, each as a ratio of an 800 MHz A9 and 3 Gbps of bandwidth. Given the lopsidedness of this ratio, the ARM core is generally indicated to be worthwhile for computation having many parallel sessions (such that the hardware's prefetching of session-specific data and TCP/reassembly offloads a large portion of the CPU load) and those requiring minimal general purpose processing of data.

Essentially zero-overhead context switching is also possible using modules as disclosed in FIG. 4-5. Because per packet processing has minimum state associated with it, and represents inherent engineered parallelism, minimal memory access is needed, aside from packet buffering. On the other hand, after packet reconstruction, the entire memory state of the session can be accessed, and so can require maximal memory utility. By using the time of packet-level processing to prefetch the next hardware scheduled application-level service context in two different processing passes, the memory can always be available for prefetching. Additionally, the FPGA 436 can hold a supplemental “ping-pong” cache (435) that is read and written with every context switch, while the other is in use. As previously noted, this is enabled in part by the SCU 432, which allows one to read out and write in memory coherently, and ACP 434 for coherent supplementation of the cache throughout the FPGA 436. The RLDRAM 438 provides for read and write to the ping-pong cache supplement 435 (show as Block1$ and Block2$) during packet-level meta-data processing. In the embodiment shown, only locally terminating queues can prompt context switching.
In operation, metadata transport code can relieve a main or host processor from tasks including fragmentation and reassembly, and checksum and other metadata services (e.g., accounting, IPSec, SSL, Overlay, etc.). As IO data streams in and out, L1 cache 437 can be filled during packet processing. During a context switch, the lock-down portion of a translation lookaside buffer (TLB) of an L1 cache can be rewritten with the addresses corresponding to the new context. In one very particular implementation, the following four commands can be executed for the current memory space.

- MRC p15,0,r0,c10,c0,0; read the lockdown register
- BIC r0,r0,#1; clear preserve bit
- MCR p15,0,r0,c10,c0,0; write to the lockdown register;
- write to the old value to the memory mapped Block RAM
  This is a small 32 cycle overhead to bear. Other TLB entries can be used by the XIMM stochastically.

Bandwidths and capacities of the memories can be precisely allocated to support context switching as well as applications such as Openflow processing, billing, accounting, and header filtering programs.
For additional performance improvements, the ACP 434 can be used not just for cache supplementation, but hardware functionality supplementation, in part by exploitation of the memory space allocation. An operand can be written to memory and the new function called, through customizing specific Open Source libraries, so putting the thread to sleep and a hardware scheduler can validate it for scheduling again once the results are ready. For example, OpenVPN uses the OpenSSL library, where the encrypt/decrypt functions 439 can be memory mapped. Large blocks are then available to be exported without delay, or consuming the L2 cache 440, using the ACP 434. Hence, a minimum number of calls are needed within the processing window of a context switch, improving overall performance.
It should be appreciated that in the foregoing description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the disclosure aiding in the understanding of one or more of the various inventive aspects. This method of disclosure, however, is not to be interpreted as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this invention.
It is also understood that the embodiments of the invention may be practiced in the absence of an element and/or step not specifically disclosed. That is, an inventive feature of the invention may be elimination of an element.
Accordingly, while the various aspects of the particular embodiments set forth herein have been described in detail, the present invention could be subject to various changes, substitutions, and alterations without departing from the spirit and scope of the invention.

Claims

What is claimed is:

1. A system, comprising:

at least one processor module connectable to a memory bus, the processor module including;

at least one memory,

at least one offload processor mounted on the processor module, and configured to execute operations on data received over the memory bus, and to output context data to the memory and read context data from the memory, and

a hardware scheduling logic mounted on the module and configured to control operations of the at least one processor.

2. The system of claim 1, further including:

an in-line module connector configured to physically connect the processor module to at least one in-line memory slot of a memory bus, with the in-line module connector being compatible with a dual-in-line-memory module (DIMM) connector.

3. The system of claim 1, wherein:

the memory bus comprises a double-data-rate (DDR) memory bus.

4. The system module of claim 1, further including:

at least one host processor coupled to the memory bus; and

the at least one processor module further includes an interface configured to receive write data and provide read data over the memory bus in response to a bus controller.

5. The system of claim 4, further including:

the interface comprises a direct memory access (DMA) slave.

6. The system of claim 4, further including:

an input/output device coupled to the at least one host processor via a host bus different than the memory bus.

7. The system of claim 6, wherein:

the input/output device is configured to transfer network packet data over the memory bus, the network packet data including a payload and a header portion.

8. The system of claim 1, further including:

an input/output memory management unit configured to translate logical read and write destinations to physical read and write destinations, the physical read and write destinations including locations of the at least one processor module.

9. The system of claim 1, further including:

at least one memory module that includes

an in-module connector configured to physically connect the memory module to one slot of the system memory bus, and

a plurality of memory integrated circuit devices configured to provide a physical memory addresses accessible via the system memory bus, including a ping-pong cache.