Application performance monitoring and management platform with anomalous flowlet resolution

Abstract

An application and network analytics platform can capture telemetry from servers and network devices operating within a network. The application and network analytics platform can determine an application dependency map (ADM) for an application executing in the network. Using the ADM, the application and network analytics platform can resolve flows into flowlets of various granularities, and determine baseline metrics for the flowlets. The baseline metrics can include transmission times, processing times, and/or data sizes for the flowlets. The application and network analytics platform can compare new flowlets against the baselines to assess availability, load, latency, and other performance metrics for the application. In some implementations, the application and network analytics platform can automate remediation of unavailability, load, latency, and other application performance issues.

Description

TECHNICAL FIELD

The subject matter of this disclosure relates in general to the field of computer networks, and more specifically for application performance monitoring and management.

BACKGROUND

An application performance management or application performance monitoring (APM) system is a tool that an enterprise can use to assess the states of applications operating in the enterprise's data center, such as availability, load, latency, and other performance issues that may arise for the applications. Some conventional networks rely on an application's built-in monitoring utilities but not every application includes these utilities. Moreover, taking such an ad-hoc approach for application monitoring may be untenable to manage for a typical enterprise network with hundreds or thousands of applications. Certain conventional APM platforms may require injecting code into an application to integrate the application with these platforms. This is a non-trivial task that can require careful coordination between the development team and the team tasked with managing the application thereafter. This can also affect the performance of the application because logging and other monitoring features occur directly in the execution path of the application. This approach is also inadequate for third party applications and/or applications deployed off-premises (e.g., via a software as a service (SaaS) or public cloud provider) that the enterprise cannot modify. Some conventional APM systems use measurement hardware and/or software on servers hosting applications while others may rely on similar tools on network devices. Both approaches can negatively affect performance of the hosting hardware and/or software. In addition, these techniques are incomplete as they neglect network performance over application performance or vice versa.

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 illustrates an example of an application and network analytics platform for providing application performance monitoring (APM) in accordance with an embodiment;

FIG. 2 illustrates an example of a forwarding pipeline of an application-specific integrated circuit (ASIC) of a network device in accordance with an embodiment;

FIG. 3 illustrates an example of an APM engine in accordance with an embodiment;

FIG. 4 illustrates an example of a network environment in accordance with an embodiment;

FIG. 5 illustrates an example of a request and response sequence in accordance with an embodiment;

FIG. 6 illustrates an example method for resolving a flow into flowlets for application performance monitoring and management in accordance with an embodiment; and

FIG. 7A and FIG. 7B illustrate examples of systems in accordance with some embodiments.

DESCRIPTION OF EXAMPLE EMBODIMENTS

Overview

An application and network analytics platform can capture telemetry (e.g., flow data, server data, process data, user data, policy data, etc.) from servers (e.g., physical and/or virtual servers) (sometimes also referred to as hosts, endpoints, computing devices, machines, and the like) and network devices (e.g., switches, routers, hubs, etc.) operating within a network. The application and network analytics platform can determine an application dependency map (ADM) for one or more applications executing in the network. Using the ADM, the application and network analytics platform can resolve flows into flowlets of varying granularities, and determine baseline metrics for the flowlets. The baseline metrics can include transmission times, processing times, and/or data sizes (in number of packets and/or bytes) for the flowlets. The application and network analytics platform can compare new flowlets against the baselines to assess availability, load, latency, and other performance metrics for one or more applications. In some embodiments, the application and network analytics platform can automate or substantially automate remediation of unavailability, load, latency, and other application performance issues.

Description

FIG. 1 illustrates an example of an application and network analytics platform 100 in accordance with an embodiment. Tetration Analytics™ provided by Cisco Systems®, Inc. of San Jose California is an example implementation of the application and network analytics platform 100. However, one skilled in the art will understand that FIG. 1 (and generally any system discussed in this disclosure) is but one possible embodiment of an application and network analytics platform and that other embodiments can include additional, fewer, or alternative components arranged in similar or alternative orders, or in parallel, unless otherwise stated. In the example of FIG. 1, the application and network analytics platform 100 includes a data collection layer 110, an analytics engine 120, and a presentation layer 140.

The data collection layer 110 may include software sensors 112, hardware sensors 114, and customer/third party data sources 116. The software sensors 112 can run within servers of a network, such as physical or bare-metal servers; hypervisors, virtual machine monitors, container orchestrators, or other virtual entity managers; virtual machines, containers, or other virtual entities. The hardware sensors 114 can reside on the application-specific integrated circuits (ASICs) of switches, routers, or other network devices (e.g., packet capture (pcap) appliances such as a standalone packet monitor, a device connected to a network device's monitoring port, a device connected in series along a main trunk of a data center, or similar device). The software sensors 112 can capture telemetry (e.g., flow data, server data, process data, user data, policy data, etc.) from the servers and the hardware sensors 114 can capture network telemetry (e.g., flow data) from network devices, and send the telemetry to the analytics engine 120 for further processing. For example, the software sensors 112 can sniff packets sent over their hosts' physical or virtual network interface cards (NICs), or individual processes on each server can report the telemetry to the software sensors 112. The hardware sensors 114 can capture network telemetry at line rate from all ports of the network devices hosting the hardware sensors.

FIG. 2 illustrates an example of a unicast forwarding pipeline 200 of an ASIC for a network device that can capture network telemetry at line rate with minimal impact on the CPU. In some embodiments, one or more network devices may incorporate the Cisco® ASE2 or ASE3 ASICs for implementing the forwarding pipeline 200. For example, certain embodiments include one or more Cisco Nexus® 9000 Series Switches provided by Cisco Systems® that utilize the ASE2 or ASE3 ASICs or equivalent ASICs. The ASICs may have multiple slices (e.g., the ASE2 and ASE3 have six slices and two slices, respectively) in which each slice represents a switching subsystem with both an ingress forwarding pipeline 210 and an egress forwarding pipeline 220. The ingress forwarding pipeline 210 can include an input/output (I/O) component, ingress MAC 212; an input forwarding controller 214; and an input data path controller 216. The egress forwarding pipeline 220 can include an output data path controller 222, an output forwarding controller 224, and an I/O component, egress MAC 226. The slices may connect to a broadcast network 230 that can provide point-to-multipoint connections from each slice and all-to-all connectivity between slices. The broadcast network 230 can provide enough bandwidth to support full-line-rate forwarding between all slices concurrently. When a packet enters a network device, the packet goes through the ingress forwarding pipeline 210 of the slice on which the port of the ingress MAC 212 resides, traverses the broadcast network 230 to get onto the egress slice, and then goes through the egress forwarding pipeline 220 of the egress slice. The input forwarding controller 214 can receive the packet from the port of the ingress MAC 212, parse the packet headers, and perform a series of lookups to determine whether to forward the packet and how to forward the packet to its intended destination. The input forwarding controller 214 can also generate instructions for the input data path controller 216 to store and queue the packet. In some embodiments, the network device may be a cut-through switch such that the network device performs input forwarding while storing the packet in a pause buffer block (not shown) of the input data path controller 216.

As discussed, the input forwarding controller 214 may perform several operations on an incoming packet, including parsing the packet header, performing an L2 lookup, performing an L3 lookup, processing an ingress access control list (ACL), classifying ingress traffic, and aggregating forwarding results. Although describing the tasks performed by the input forwarding controller 214 in this sequence, one of ordinary skill will understand that, for any process discussed herein, there can be additional, fewer, or alternative steps performed in similar or alternative orders, or in parallel, within the scope of the various embodiments unless otherwise stated.

In some embodiments, when a unicast packet enters through a front-panel port (e.g., a port of ingress MAC 212), the input forwarding controller 214 may first perform packet header parsing. For example, the input forwarding controller 214 may parse the first 128 bytes of the packet to extract and save information such as the L2 header, EtherType, L3 header, and TCP IP protocols.

As the packet goes through the ingress forwarding pipeline 210, the packet may be subject to L2 switching and L3 routing lookups. The input forwarding controller 214 may first examine the destination MAC address of the packet to determine whether to switch the packet (i.e., L2 lookup) or route the packet (i.e., L3 lookup). For example, if the destination MAC address matches the network device's own MAC address, the input forwarding controller 214 can perform an L3 routing lookup. If the destination MAC address does not match the network device's MAC address, the input forwarding controller 214 may perform an L2 switching lookup based on the destination MAC address to determine a virtual LAN (VLAN) identifier. If the input forwarding controller 214 finds a match in the MAC address table, the input forwarding controller 214 can send the packet to the egress port. If there is no match for the destination MAC address and VLAN identifier, the input forwarding controller 214 can forward the packet to all ports in the same VLAN.

During L3 routing lookup, the input forwarding controller 214 can use the destination IP address for searches in an L3 host table. This table can store forwarding entries for directly attached hosts and learned/32 host routes. If the destination IP address matches an entry in the host table, the entry will provide the destination port, next-hop MAC address, and egress VLAN. If the input forwarding controller 214 finds no match for the destination IP address in the host table, the input forwarding controller 214 can perform a longest-prefix match (LPM) lookup in an LPM routing table.

In addition to forwarding lookup, the input forwarding controller 214 may also perform ingress ACL processing on the packet. For example, the input forwarding controller 214 may check ACL ternary content-addressable memory (TCAM) for ingress ACL matches. In some embodiments, each ASIC may have an ingress ACL TCAM table of 4000 entries per slice to support system internal ACLs and user-defined ingress ACLs. These ACLs can include port ACLs, routed ACLs, and VLAN ACLs, among others. In some embodiments, the input forwarding controller 214 may localize the ACL entries per slice and program them only where needed.

In some embodiments, the input forwarding controller 214 may also support ingress traffic classification. For example, from an ingress interface, the input forwarding controller 214 may classify traffic based on the address field, IEEE 802.1q class of service (CoS), and IP precedence or differentiated services code point in the packet header. In some embodiments, the input forwarding controller 214 can assign traffic to one of eight quality-of-service (QoS) groups. The QoS groups may internally identify the traffic classes used for subsequent QoS processes as packets traverse the system.

In some embodiments, the input forwarding controller 214 may collect the forwarding metadata generated earlier in the pipeline (e.g., during packet header parsing, L2 lookup, L3 lookup, ingress ACL processing, ingress traffic classification, forwarding results generation, etc.) and pass it downstream through the input data path controller 216. For example, the input forwarding controller 214 can store a 64-byte internal header along with the packet in the packet buffer. This internal header can include 16 bytes of iETH (internal communication protocol) header information, which the input forwarding controller 214 can prepend to the packet when transferring the packet to the output data path controller 222 through the broadcast network 230. The network device can strip the 16-byte iETH header when the packet exits the front-panel port of the egress MAC 226. The network device may use the remaining internal header space (e.g., 48 bytes) to pass metadata from the input forwarding queue to the output forwarding queue for consumption by the output forwarding engine.

In some embodiments, the input data path controller 216 can perform ingress accounting functions, admission functions, and flow control for a no-drop class of service. The ingress admission control mechanism can determine whether to admit the packet into memory based on the amount of buffer memory available and the amount of buffer space already used by the ingress port and traffic class. The input data path controller 216 can forward the packet to the output data path controller 222 through the broadcast network 230.

As discussed, in some embodiments, the broadcast network 230 can comprise a set of point-to-multipoint wires that provide connectivity between all slices of the ASIC. The input data path controller 216 may have a point-to-multipoint connection to the output data path controller 222 on all slices of the network device, including its own slice.

In some embodiments, the output data path controller 222 can perform egress buffer accounting, packet queuing, scheduling, and multicast replication. In some embodiments, all ports can dynamically share the egress buffer resource. In some embodiments, the output data path controller 222 can also perform packet shaping. In some embodiments, the network device can implement a simple egress queuing architecture. For example, in the event of egress port congestion, the output data path controller 222 can directly queue packets in the buffer of the egress slice. In some embodiments, there may be no virtual output queues (VoQs) on the ingress slice. This approach can simplify system buffer management and queuing.

As discussed, in some embodiments, one or more network devices can support up to 10 traffic classes on egress, 8 user-defined classes identified by QoS group identifiers, a CPU control traffic class, and a switched port analyzer (SPAN) traffic class. Each user-defined class can have a unicast queue and a multicast queue per egress port. This approach can help ensure that no single port will consume more than its fair share of the buffer memory and cause buffer starvation for other ports.

In some embodiments, multicast packets may go through similar ingress and egress forwarding pipelines as the unicast packets but instead use multicast tables for multicast forwarding. In addition, multicast packets may go through a multistage replication process for forwarding to multiple destination ports. In some embodiments, the ASIC can include multiple slices interconnected by a non-blocking internal broadcast network. When a multicast packet arrives at a front-panel port, the ASIC can perform a forwarding lookup. This lookup can resolve local receiving ports on the same slice as the ingress port and provide a list of intended receiving slices that have receiving ports in the destination multicast group. The forwarding engine may replicate the packet on the local ports, and send one copy of the packet to the internal broadcast network, with the bit vector in the internal header set to indicate the intended receiving slices. In this manner, only the intended receiving slices may accept the packet off of the wire of the broadcast network. The slices without receiving ports for this group can discard the packet. The receiving slice can then perform local L3 replication or L2 fan-out lookup and replication to forward a copy of the packet to each of its local receiving ports.

In FIG. 2, the forwarding pipeline 200 also includes a flow cache 240, which when combined with direct export of collected telemetry from the ASIC (i.e., data hardware streaming), can enable collection of packet and flow metadata at line rate while avoiding CPU bottleneck or overhead. The flow cache 240 can provide a full view of packets and flows sent and received by the network device. The flow cache 240 can collect information on a per-packet basis, without sampling and without increasing latency or degrading performance of the network device. To accomplish this, the flow cache 240 can pull information from the forwarding pipeline 200 without being in the traffic path (i.e., the ingress forwarding pipeline 210 and the egress forwarding pipeline 220).

In addition to the traditional forwarding information, the flow cache 240 can also collect other metadata such as detailed IP and TCP flags and tunnel endpoint identifiers. In some embodiments, the flow cache 240 can also detect anomalies in the packet flow such as inconsistent TCP flags. The flow cache 240 may also track flow performance information such as the burst and latency of a flow. By providing this level of information, the flow cache 240 can produce a better view of the health of a flow. Moreover, because the flow cache 240 does not perform sampling, the flow cache 240 can provide complete visibility into the flow.

In some embodiments, the flow cache 240 can include an events mechanism to complement anomaly detection. This configurable mechanism can define a set of parameters that represent a packet of interest. When a packet matches these parameters, the events mechanism can trigger an event on the metadata that triggered the event (and not just the accumulated flow information). This capability can give the flow cache 240 insight into the accumulated flow information as well as visibility into particular events of interest. In this manner, networks, such as a network implementing the application and network analytics platform 100, can capture telemetry more comprehensively and not impact application and network performance.

Returning to FIG. 1, the telemetry captured by the software sensors 112 and hardware sensors 114 can include metadata relating to individual packets (e.g., packet size, source address, source port, destination address, destination port, etc.); flows (e.g., number of packets and aggregate size of packets having the same source address/port, destination address/port, L3 protocol type, class of service, router/switch interface, etc. sent/received without inactivity for a certain time (e.g., 15 seconds) or sent/received over a certain duration (e.g., 30 minutes)); flowlets (e.g., flows of sub-requests and sub-responses generated as part of an original request or response flow and sub-flows of these flows); bidirectional flows (e.g., flow data for a request/response pair of flows having corresponding source address/port, destination address/port, etc.); groups of flows (e.g., flow data for flows associated with a certain process or application, server, user, etc.), sessions (e.g., flow data for a TCP session); or other types of network communications of specified granularity. That is, the network telemetry can generally include any information describing communication on all layers of the Open Systems Interconnection (OSI) model. In some embodiments, the network telemetry collected by the sensors 112 and 114 can also include other network traffic data such as hop latency, packet drop count, port utilization, buffer information (e.g., instantaneous queue length, average queue length, congestion status, etc.), and other network statistics.

In some embodiments, the application and network analytics platform 100 can resolve flows into flowlets (i.e., sequences of requests and responses of a larger request and response) of various granularities. For example, a response to a request to an enterprise application may result in multiple sub-requests and sub-responses to various back-end services (e.g., authentication, static content, data, search, sync, etc.). The application and network analytics platform 100 can break a flow into its constituent components to provide greater insight into application and network performance. The application and network analytics platform 100 can perform this resolution in real time or substantially real time (e.g., no more than a few minutes after detecting the flow).

The application and network analytics platform 100 can associate a flow with a server sending or receiving the flow, an application or process triggering the flow, the owner of the application or process, and one or more policies applicable to the flow, among other telemetry. The telemetry captured by the software sensors 112 can thus include server data, process data, user data, policy data, and other data (e.g., virtualization information, tenant information, sensor information, etc.). The server telemetry can include the server name, network address, CPU usage, network usage, disk space, ports, logged users, scheduled jobs, open files, and similar information. In some embodiments, the server telemetry can also include information about the file system of the server, such as the lists of files (e.g., log files, configuration files, device special files, etc.) and/or directories stored within the file system as well as the metadata for the files and directories (e.g., presence, absence, or modifications of a file and/or directory). In some embodiments, the server telemetry can further include physical or virtual configuration information (e.g., processor type, amount of random access memory (RAM), amount of disk or storage, type of storage, system type (e.g., 32-bit or 64-bit), operating system, public cloud provider, virtualization platform, etc.).

The process telemetry can include the process name (e.g., bash, httpd, netstat, etc.), process identifier, parent process identifier, path to the process (e.g., /usr2/username/bin/, /usr/local/bin,/usr/bin, etc.), CPU utilization, memory utilization, memory address, scheduling information, nice value, flags, priority, status, start time, terminal type, CPU time taken by the process, and the command string that initiated the process (e.g., “/opt/tetration/collector/tet-collector--config_file/etc/tetration/collector/collector.config--timest amp_flow_info--logtostderr--utc_time_in_file_name true--max_num_ssl_sw_sensors 63000--enable_client_certificate true”). The user telemetry can include information regarding a process owner, such as the user name, user identifier, user's real name, e-mail address, user's groups, terminal information, login time, expiration date of login, idle time, and information regarding files and/or directories of the user.

The customer/third party data sources 116 can include out-of-band data such as power level, temperature, and physical location (e.g., room, row, rack, cage door position, etc.). The customer/third party data sources 116 can also include third party data regarding a server such as whether the server is on an IP watch list or security report (e.g., provided by Cisco®, Arbor Networks® of Burlington, Massachusetts, Symantec® Corp. of Sunnyvale, California, Sophos® Group plc of Abingdon, England, Microsoft® Corp. of Seattle, Washington, Verizon® Communications, Inc. of New York, New York, among others), geolocation data, and Whois data, and other data from external sources.

In some embodiments, the customer/third party data sources 116 can include data from a configuration management database (CMDB) or configuration management system (CMS) as a service. The CMDB/CMS may transmit configuration data in a suitable format (e.g., JavaScript® object notation (JSON), extensible mark-up language (XML), yet another mark-up language (YAML), etc.)).

The processing pipeline 122 of the analytics engine 120 can collect and process the telemetry. In some embodiments, the processing pipeline 122 can retrieve telemetry from the software sensors 112 and the hardware sensors 114 every 100 ms or faster. Thus, the application and network analytics platform 100 may not miss or is much less likely than conventional systems (which typically collect telemetry every 60 seconds) to miss “mouse” flows. In addition, as the telemetry tables flush so often, the software sensors 112 and the hardware sensors 114 do not or are much less likely than conventional systems to drop telemetry because of overflow/lack of memory. An additional advantage of this approach is that the application and network analytics platform is responsible for flow-state tracking instead of network devices. Thus, the ASICs of the network devices of various embodiments can be simpler or can incorporate other features.

In some embodiments, the processing pipeline 122 can filter out extraneous or duplicative data or it can create summaries of the telemetry. In some embodiments, the processing pipeline 122 may process (and/or the software sensors 112 and hardware sensors 114 may capture) only certain types of telemetry and disregard the rest. For example, the processing pipeline 122 may process (and/or the sensors may monitor) only high-priority telemetry, telemetry associated with a particular subnet (e.g., finance department, human resources department, etc.), telemetry associated with a particular application (e.g., business-critical applications, compliance software, health care applications, etc.), telemetry from external-facing servers, etc. As another example, the processing pipeline 122 may process (and/or the sensors may capture) only a representative sample of telemetry (e.g., every 1,000th packet or other suitable sample rate).

Collecting and/or processing telemetry from multiple servers of the network (including within multiple partitions of virtualized hosts) and from multiple network devices operating between the servers can provide a comprehensive view of network behavior. The capture and/or processing of telemetry from multiple perspectives rather than just at a single device located in the data path (or in communication with a component in the data path) can allow the data to be correlated from the various data sources, which may be used as additional data points by the analytics engine 120.

In addition, collecting and/or processing telemetry from multiple points of view can enable capture of more accurate data. For example, a conventional network may consist of external-facing network devices (e.g., routers, switches, network appliances, etc.) such that the conventional network may not be capable of monitoring east-west traffic, including telemetry for VM-to-VM or container-to-container communications on a same host. As another example, the conventional network may drop some packets before those packets traverse a network device incorporating a sensor. The processing pipeline 122 can substantially mitigate or eliminate these issues altogether by capturing and processing telemetry from multiple points of potential failure. Moreover, the processing pipeline 122 can verify multiple instances of data for a flow (e.g., telemetry from a source (physical server, hypervisor, container orchestrator, other virtual entity manager, VM, container, and/or other virtual entity), one or more network devices, and a destination) against one another.

In some embodiments, the processing pipeline 122 can assess a degree of accuracy of telemetry for a single flow captured by multiple sensors and utilize the telemetry from a single sensor determined to be the most accurate and/or complete. The degree of accuracy can be based on factors such as network topology (e.g., a sensor closer to the source may be more likely to be more accurate than a sensor closer to the destination), a state of a sensor or a server hosting the sensor (e.g., a compromised sensor/server may have less accurate telemetry than an uncompromised sensor/server), or telemetry volume (e.g., a sensor capturing a greater amount of telemetry may be more accurate than a sensor capturing a smaller amount of telemetry).

In some embodiments, the processing pipeline 122 can assemble the most accurate telemetry from multiple sensors. For instance, a first sensor along a data path may capture data for a first packet of a flow but may be missing data for a second packet of the flow while the reverse situation may occur for a second sensor along the data path. The processing pipeline 122 can assemble data for the flow from the first packet captured by the first sensor and the second packet captured by the second sensor.

In some embodiments, the processing pipeline 122 can also disassemble or decompose a flow into sequences of request and response flowlets (e.g., sequences of requests and responses of a larger request or response) of various granularities. For example, a response to a request to an enterprise application may result in multiple sub-requests and sub-responses to various back-end services (e.g., authentication, static content, data, search, sync, etc.). The processing pipeline 122 can break a flow down into its constituent components to provide greater insight into application and network performance. The processing pipeline 122 can perform this resolution in real time or substantially real time (e.g., no more than a few minutes after detecting the flow).

The processing pipeline 122 can store the telemetry in a data lake (not shown), a large-scale storage repository characterized by massive storage for various types of data, enormous processing power, and the ability to handle nearly limitless concurrent tasks or jobs. In some embodiments, the analytics engine 120 may deploy at least a portion of the data lake using the Hadoop® Distributed File System (HDFS™) from Apache® Software Foundation of Forest Hill, Maryland HDFS™ is a highly scalable and distributed file system that can scale to thousands of cluster nodes, millions of files, and petabytes of data. A feature of HDFS™ is its optimization for batch processing, such as by coordinating data computation to where data is located. Another feature of HDFS™ is its utilization of a single namespace for an entire cluster to allow for data coherency in a write-once, read-many access model. A typical HDFS™ implementation separates files into blocks, which are typically 64 MB in size and replicated in multiple data nodes. Clients can access data directly from the data nodes.

The processing pipeline 122 can propagate the processed data to one or more engines, monitors, and other components of the analytics engine 120 (and/or the components can retrieve the data from the data lake), such as an application dependency mapping (ADM) engine 124, an inventory monitor 126, a flow monitor 128, and an application performance monitoring (APM) engine 130.

The ADM engine 124 can determine dependencies of applications running in the network, i.e., how processes on different servers interact with one another to perform the functions of the application. Particular patterns of traffic may correlate with particular applications. The ADM engine 124 can evaluate telemetry processed by the processing pipeline 122 to determine the interconnectivity or dependencies of the application to generate a graph for the application (i.e., an application dependency mapping). For example, in a conventional three-tier architecture for a web application, first servers of the web tier, second servers of the application tier, and third servers of the data tier make up the web application. From flow data, the ADM engine 124 may determine that there is first traffic flowing between external servers on port 80 of the first servers corresponding to Hypertext Transfer Protocol (HTTP) requests and responses. The flow data may also indicate second traffic between first ports of the first servers and second ports of the second servers corresponding to application server requests and responses and third traffic flowing between third ports of the second servers and fourth ports of the third servers corresponding to database requests and responses. The ADM engine 124 may define an application dependency map or graph for this application as a three-tier application including a first endpoint group (EPG) (i.e., groupings of application tiers or clusters, applications, and/or application components for implementing forwarding and policy logic) comprising the first servers, a second EPG comprising the second servers, and a third EPG comprising the third servers.

The inventory monitor 126 can continuously track the network's assets (e.g., servers, network devices, applications, etc.) based on the telemetry processed by the processing pipeline 122. In some embodiments, the inventory monitor 126 can assess the state of the network at a specified interval (e.g., every 1 minute). That is, the inventory monitor 126 can periodically take snapshots of the states of applications, servers, network devices, and/or other elements of the network. In other embodiments, the inventory monitor 126 can capture the snapshots when events of interest occur, such as an application experiencing latency that exceeds an application latency threshold; the network experiencing latency that exceeds a network latency threshold; failure of a server, network device, or other network element; and similar circumstances. Snapshots can include a variety of telemetry associated with network elements. For example, a snapshot of a server can represent the processes executing on the server at a time of capture, the amount of CPU utilized by each process (e.g., as an amount of time and/or a relative percentage), the amount of memory utilized by each process (e.g., in bytes and/or as a relative percentage), the amount of disk utilized by each process (e.g., in bytes or as a relative percentage), and a distance (physical and/or logical, relative and/or absolute) from one or more other network elements.

In some embodiments, on a change to the network (e.g., a server updating its operating system or running a new process; a server communicating on a new port; a VM, container, or other virtualized entity migrating to a different host and/or subnet, VLAN, VxLAN, or other network segment; etc.), the inventory monitor 126 can alert the APM engine 130 to ensure that applications and the network remain performing as expected in view of the change(s) to the data center.

The flow monitor 128 can analyze flows to detect whether they are associated with anomalous or malicious traffic. In some embodiments, the flow monitor 128 may receive examples of past flows determined to perform at expectation (i.e., the length of time for the flow to reach a network device or a destination is within a threshold length of time) or perform below expectation (i.e., the length of time for the flow to reach the network device and/or destination exceeds the threshold length of time). The flow monitor 128 can utilize machine learning to analyze the telemetry processed by the processing pipeline 122 and classify each current flow based on similarity to past flows. On detection of an anomalous flow, such as a flow taking a shorter or longer duration from source to destination than a specified time range, a flow of a size less or more than a specified amount, or a flow previously classified as a network attack, the flow monitor 128 may transmit an alert to the APM engine 130 and/or to the presentation layer 140. In some embodiments, the network may operate within a trusted environment for a period of time so that the analytics engine 120 can establish a baseline of normal operation.

The APM engine 130 can be responsible for monitoring and managing applications running within the network. For example, the APM engine 130 may receive an alert from the inventory monitor 126 on a change to the network or an alert from the flow monitor upon the flow monitor 128 detecting an anomalous flow. The APM engine 130 can evaluate the telemetry to pinpoint the root cause of degraded performance. In some embodiments, the APM engine 130 can also perform remediation tasks, such as load balancing, migrating application components to be closer together (physically and/or logically), instantiating new server instances (in the private network and/or a public cloud), and disabling network connectivity for problematic servers, among other operations.

FIG. 3 illustrates an example of an application performance monitoring (APM) engine 300 that represents one of many possible implementations of the APM engine 130. The APM engine 300 can include one or more performance front end processes (PFEs) 310, a coordinator cluster 320, and a telemetry store 330. While the APM engine 300 includes specific components in this example, one of ordinary skill in the art will understand that the configuration of the APM engine 300 is one possible configuration and that other configurations with more or less components are also possible.

FIG. 3 shows the PFEs 310 in communication with sensors 302. The sensors 302 represent one of many possible implementations of the software sensors 112 and/or hardware sensors 114 of FIG. 1. After installation on a server and/or network device of the data center, each sensor 302 can register with the coordinator cluster 320 via communication with one or more of the PFEs 310. Upon successful registration, the sensor 302 may begin capturing telemetry, including flow data, host data, process data, user data, policy data, etc. The sensors 302 may receive new configuration information from time to time, such as to capture additional telemetry for diagnostic purposes or for software upgrades. In some embodiments, the APM engine 300 may encode the configuration information in a high-level, platform-independent format. In some embodiments, each sensor 302 can determine its server's operating environment, convert the high-level configuration information into platform-specific configuration information, and apply certain platform-specific optimizations based on the operating environment. In other embodiments, the APM engine 300 may translate the high-level configuration information to the platform-specific format remotely from the sensors 302 before distribution.

In some embodiments, the PFEs 310 can be responsible for storing platform-independent configuration information in memory, handling registration of the sensors 302, monitoring updates to the configuration information, distributing the updates to the sensors 302, and collecting telemetry captured by the sensors 302. In the example of FIG. 3, the PFEs 310 can function as intermediaries between the sensors 302 and the coordinator cluster 320. This can add a layer of security between servers and the APM engine 300. For example, the sensors 302 can operate under the least-privileged principle having trust in only the coordinator cluster 320 and no trust in the PFEs 310. The sensors 302 and the PFEs 310 must sign and authenticate all transactions between them, including configuration and registration.

The coordinator cluster 320 can operate as the controller for the APM engine 300. In the example of FIG. 3, the coordinator cluster 320 implements a high availability scheme (e.g., ZooKeeper, doozerd, and etcd) in which the cluster elects one coordinator instance master and the remaining coordinator instances serve as standby instances. The coordinator cluster 320 can manage the assignment of the sensors 302 to the PFEs 310. In some embodiments, each sensor 302 may initially register with the PFE 310 closest (physically and/or logically) to the sensor's host but the coordinator cluster 320 may reassign the sensor to a different PFE, such as for load balancing and/or in the event of the failure of one or more of the PFEs 310. In some embodiments, the coordinator cluster 320 may use sharding for load balancing and providing high availability for the PFEs 310.

The telemetry store 330 can maintain sensor data captured by the sensors 302. In some embodiments, the APM engine can maintain recently captured and/or accessed telemetry in more readily-accessible data stores (e.g., solid state devices (SSD), optimized hard disk drives (HDD), etc.) and migrate older telemetry to less accessible data stores (e.g., commodity HDDs, tape, etc.). In some embodiments, the APM engine 300 may implement the telemetry store 330 using Druid® or other relational database platform. In other embodiments, the APM engine 300 may implement the telemetry store 330 using software provided by MongoDB®, Inc. of New York, New York or other NoSQL database.

In some embodiments, the coordinator cluster 320 may also be responsible for load balancing the PFEs 310, ensuring high availability of the PFEs 310 to the sensors 302, and receiving and storing telemetry in the telemetry store 330. In other embodiments, the APM engine 300 can integrate the functionality of a PFE and a coordinator or further divide the functionality of the PFE and the coordinator into additional components.

Returning to FIG. 1, the presentation layer 140 can include a web graphical user interface (GUI) 142, API endpoints 144, and an event-based notification system 146. In some embodiments, the application and network analytics platform 100 may implement the web GUI 142 using Ruby on Rails™ as the web application framework. Ruby on Rails™ is model-view-controller (MVC) framework that provides default structures for a database, a web service, and web pages. Ruby on Rails™ relies on web standards such as JSON or XML for data transfer, and hypertext markup language (HTML), cascading style sheets, (CSS), and JavaScript® for display and user interfacing.

In some embodiments, the application and network analytics platform 100 can expose application programming interface (API) endpoints (e.g., such as those based on the simple object access protocol (SOAP), a service oriented architecture (SOA), a representational state transfer (REST) architecture, a resource oriented architecture (ROA), etc.) for monitor the performance of applications executing in a network and the network itself. In some embodiments, the application and network analytics platform 100 may implement the API endpoints 144 using Hadoop® Hive from Apache® for the back end, and Java® Database Connectivity (JDBC) from Oracle® Corporation of Redwood Shores, California, as an API layer. Hive is a data warehouse infrastructure that provides data summarization and ad hoc querying. Hive provides a mechanism to query data using a variation of structured query language (SQL) called HiveQL. JDBC is an application programming interface (API) for the programming language Java®, which defines how a client may access a database.

In some embodiments, the application and network analytics platform 100 may implement the event-based notification system using Hadoop® Kafka. Kafka is a distributed messaging system that supports partitioning and replication. Kafka uses the concept of topics. Topics are feeds of messages in specific categories. In some embodiments, Kafka can take raw packet captures and telemetry information as input, and output messages to a security information and event management (SIEM) platform that provides users with the capability to search, monitor, and analyze machine-generated data.

In some embodiments, each server in the network may include a software sensor and each network device may include a hardware sensor 114. In other embodiments, the software sensors 112 and hardware sensors 114 can reside on a portion of the servers and network devices of the network. In some embodiments, the software sensors 112 and/or hardware sensors 114 may operate in a full-visibility mode in which the sensors collect telemetry from every packet and every flow or a limited-visibility mode in which the sensors provide only the conversation view required for application insight and policy generation.

FIG. 4 illustrates an example of a network environment 400. One of ordinary skill will appreciate that, for the network environment 400 and any environment discussed in this disclosure, there can be additional or fewer nodes, devices, links, networks, or components in similar or alternative configurations. Other embodiments may have different numbers and/or types of clients, networks, nodes, cloud components, servers, software components, devices, virtual or physical resources, configurations, topologies, services, appliances, deployments, or network devices. Further, the network environment 400 can include any number or type of resources accessible and utilized by clients or tenants.

The network environment 400 can include a client computing device 402, a wide area network (WAN) 404, and a local area network (LAN) 410. Although not shown here for purposes of simplicity and conciseness, a typical data center may also include a firewall, a load balancer, and/or an additional edge router between an edge network device 406b and the web server 412. The client 402 can be any kind of computing device (i.e., of varying types, capabilities, operating systems, etc.) capable of communication over a network, such as a server (physical or virtual), a desktop computer, a laptop, a tablet, a smartphone, or a wearable device (e.g., a watch; eyeglasses, a visor, a head-mounted display or other device generally worn over a user's eyes; headphones, ear buds, or other device generally worn in or over a user's ears; etc.). The client 402 can also be an “infotainment system” (i.e., a computing device integrated with a means of transportation), a “smart” home device or Internet of Things (IoT) device (e.g., a television, a set-top box, a digital video recorder (DVR), a digital video disc (DVD) player or other media player, a video game console, etc.), or other electronic devices.

The WAN 404 can include one or more networks and/or network devices, such as the network devices 406a and 406b, for interconnecting the client 402 and the LAN 410. WANs can connect geographically dispersed nodes over long-distance communications links, such as common carrier telephone lines, optical light paths, synchronous optical networks (SONET), or synchronous digital hierarchy (SDH) links. LANs and WANs can include L2 and/or L3 networks and servers. The Internet is an example of a WAN that connects disparate networks throughout the world, providing global communication between nodes on various networks. The nodes typically communicate over the network by exchanging discrete frames or packets of data according to predefined protocols, such as the Transmission Control Protocol/Internet Protocol (TCP/IP). In this context, a protocol can refer to a set of rules defining how the nodes interact with each other. The WAN 404 can also be a private network, such as a global enterprise network, that operates using similar or the same technologies as the public Internet.

LANs can connect nodes over dedicated private communications links located in the same general physical location, such as a building or campus. In the example of FIG. 4, the LAN 410 can be a data center that includes a single, simple web application. A typical data center can include hundreds or thousands of simpler or more complex applications. The web application includes a web server 412 connected to a network device 414 (e.g., a hub, switch, router, etc.) for communicating with an authentication server 416, a content server 418, and a data server 420. The data server 420 can directly or indirectly connect to a database 422. Although FIG. 4 may appear to show that the web server 412, the authentication server 416, the content server 418, and the data server 420 as separate elements, the servers or various combinations of the servers can reside on a single physical machine as virtual machines, containers, or other virtual entities. Similarly, the network device 414 can be one or more physical network devices, virtual network devices, or various combinations of physical and virtual network devices in a variety of configurations (e.g., leaf-spine, three-tier (i.e., including core, aggregation, and access levels), fat tree, mesh, bus, hub and spoke, etc.).

In the example of FIG. 4, the web application can be a simple website associated with one or more user accounts. Users may login to their accounts and access their personal information and other content. A user may access her account associated with the website/web application from the client device 402 (e.g., via a standalone client application, web browser, plug-in, etc.). The client device 402 may communicate with the website/web application located in the LAN 410 over the WAN 404 (e.g., the Internet, an enterprise network, etc.) for the user to interact with her account in any number of ways, such as updating privileged information (e.g., name, email address, login password, etc.), viewing content, creating content, etc. Each interaction can constitute a request to the website/web application and the results of the request may be a response. However, each request may comprise one or more additional sub-requests and sub-responses. Likewise, each response can include one or more sub-requests and sub-responses. The number of sub-requests and sub-responses can largely depend on the complexity of the application. But even a simple web application/website such as illustrated in FIG. 4 can include numerous sub-requests and sub-responses.

FIG. 5 illustrates an example of a time sequence 500 of a request and response for an interaction with the website/web application of FIG. 4. One of ordinary skill will understood that, for any sequence discussed herein, there can be additional, fewer, or alternative steps performed in similar or alternative orders, or in parallel, within the scope of the various embodiments unless otherwise stated. The user interaction may be a user logging into the website/web application, clicking to view certain content, searching for content in the website/web application, among any number of possibilities. In the time sequence 500, the vertical axis corresponds to time and the horizontal axis shows the communications between the various components of the website/web application.

In the example of FIG. 5, the client 402 may send an originating request (i.e., at point A) for a particular page of the website/web application or other content. After traversing the WAN 404, the originating request may arrive at the web server 412 after a duration AB. As discussed above, the originating request may include segments through the WAN 404, one or more of the data center's edge network devices, firewalls, load balancers, among other software and hardware components but FIG. 5 excludes those segments for brevity and clarity. The web server 412 may spend some time BC processing the request to determine that the request is a request for personal content requiring authentication. Thus, the web server 412 may send an authentication request (i.e., at point C) to the authentication server 416 to verify the user's login name and password match the user's information stored by the website/web application. To send the authentication request, the web server 412 may first forward the authentication request to the network device 414 for a duration CD. The network device 414 can receive the authentication request, take some amount of time DE to process the packets of the authentication request and forward the request to the authentication server 416 for a duration EF. The authentication server 416 may process the authentication request and prepare an authentication response (i.e., at point G) for a time FG and forward the authentication request to the network device 414 for a time GH. The network device 414 can spend a time HI to determine where to forward the authentication request to the web server 412 upon which the network device 414 may forward the response to the web server 412 over a duration IJ.

The web server 412 may confirm authentication of the user from the authentication response and begin retrieving content to provide a response to the originating request. The response may be a page of the website/web application that includes content from the content server 418 and personal information from the data server 420. The web server 412 may take up a time JK to prepare the request to the content server 418 and a time KS to prepare the request to the data server 420. The web server 412 may send the content request to the network device 414 over a time KL, upon which the network device 414 may spend a time LM to forward the request to the content server 418 for a duration MN. The content server 418 can receive the content request, take a time NO to process the request, and transmit a content response to the network device 414 over a time OP. The network device 414 can process the content response for a time PQ and forward the content response to the web server after which a time QR elapses.

In parallel or very near in time to the request/response to the content server 418, the web server 412 may send the data request to the network device 414 for a time ST. The network device can process the data request for a time TU and forward the data request over a period of time UV. The data server 420 may have an architecture such that it must retrieve requested data from the database 422, and therefore must perform some time VW processing the data request and preparing a database request (i.e., at point W). The database request may take a time WX to arrive at the database 422. Fetching the requested data may occur over a duration XY and transmitting the data (i.e., at point Y) back to the data server 420 may occur over a time YZ. The data server may process the database response within a time Zα before sending a data response back to the network device 414 over a time αβ. The network device may process the data response for a time βγ and forward the data response over a time γδ. The web server 412 may assemble the content retrieved from the content server 418 and the data retrieved from the data server 420 over a time δε before sending the response to the originating request to the client 402 over a time εζ.

In the example of FIG. 5, the segments beginning from BC through εζ can represent the total latency for sending the web server response to the originating request to the web server. The total response latency can thus include latency due to the web server 412, latency due to the authentication server 416, and the maximum between the latency due to the content server 418 and the latency due to the data server 420 (assuming parallel or substantially parallel requests to the content server 418 and the data server 420). Some conventional data centers may be able to derive some of these latencies. For example, a data center relying on network telemetry from network devices (e.g., the network device 414 may be able to calculate the web server latency via the timestamps of the authentication request flowlet and the authentication response flowlet. Likewise, it may be possible for data centers that rely on telemetry from servers to compute the web server latency from web server logs indicating when the web server 412 sent the authentication request flowlet to the authentication server 416 and when the web server received the authentication response flowlet from the authentication server 416. Various embodiments can improve upon these conventional approaches by computing more granular latencies. For example, the application and network analytics platform 100 can resolve the total latency for providing the response to the authentication request to include the segment CD representing network latency from the web server 412 to the network device 414, the segment DE representing network device latency (although negligible and ignored in some embodiments), the segment EF representing network latency from the network device 414 to the authentication server 416, the segment FG represent authentication server latency, the segment GH representing network latency from the authentication server 416 to the network device 414, the segment HI representing network device latency, and the segment IJ representing the network latency from the network device 414 to the web server 412. The conventional data centers could not determine latency at this level of granularity but various embodiments of the present technology can by obtaining telemetry from the web server 412, the network device 414, and the authentication server 416. Over a period of time, an application and analytics framework in accordance with some embodiments may establish baseline metrics for the network latency between the web server 412 and the authentication server 416 (i.e., segments CD, (sometimes DE), EF), authentication server latency (i.e., segment FG), and the network latency between the authentication server 416 and the web server 412 (i.e., segments GH, (sometimes HI), and IJ). When users experience latency, an administrator may determine quickly whether the issue is due to server latency or network latency based on this manner of decomposing flows into flowlets.

As shown in FIG. 5, a single user interaction with a simple application can comprise more than a dozen requests and responses. A typical enterprise network can include hundreds or thousands of much more complex applications with hundreds or thousands of users possibly performing millions or billions of operations per day. It would not be possible for a human user to process and analyze this volume of data. FIG. 5 also demonstrates the incompleteness of data centers that only rely on telemetry from one of servers or network devices. For example, in a first data center in which telemetry is only available from the network device 414 and in a situation where the network device's attempt to communicate with the data server 420 is unusually slow or fails, the first APM platform could not diagnose whether the issue arises due to application unavailability or load. In addition, the first data center may not be able to provide insight as to whether the problem lays in the data server 420, the database 422, or the connection between the data server and the database. For instance, in the data request, the first data center may only recover ST and UV but may not be capable of recovering VW, WX, XY, YZ, and Zα because the first data center may not be aware that the data server 320 has a dependency on the database 322.

A second data center that relies only on telemetry from servers may also suffer from various defects. Such a system may be able to detect anomalous (or no) response times in communications between the web server 412 and the data server 420 but may not be able to ascertain whether latency or failure is due to the web server 412, the network device 414, the data server 420, the database 422, or the connection between the data server and the database. For example, in the data request, the second data center may have a view of the data request and the data response but ST, TU, UV, VW, WX, XY, YZ, Zα, αβ, βγ, and γδ may be a black box. Various embodiments may overcome these and other deficiencies with prior art data center by capturing telemetry along multiple points of a data path, resolving flows into flowlets, establishing baselines for the flowlets, and evaluating new traffic against the baselines to more precisely diagnose the root cause of application and network performance degradation.

FIG. 6 illustrates an example of a method 600 for resolving one or more flows into flowlets for application performance monitoring and management. One of ordinary skill will understood that, for any method discussed herein, there can be additional, fewer, or alternative steps performed in similar or alternative orders, or in parallel, within the scope of the various embodiments unless otherwise stated. A network, and particularly, an application and network analytics platform (e.g., the application and network analytics platform 100 of FIG. 1), an analytics engine (e.g., the analytics engine 120 of FIG. 1), an APM engine (e.g., the APM engine 130 of FIG. 1), a network operating system, a virtual entity manager, or similar system can perform the method 600.

In the example of FIG. 6, the method 600 may begin at step 602 in which sensors (e.g., the software sensors 112 and hardware sensors 114 of FIG. 1) capture telemetry from servers and network devices of the network (e.g., flow data, host data, process data, user data, policy data, etc.) over a first period of time to establish a baseline for application and network performance. In some embodiments, the application and network analytics platform may also collect virtualization information, network topology information, and application information (e.g., configuration information, previously generated application dependency maps, application policies, etc.). In addition, the application and network analytics platform may also collect out-of-band data (e.g., power level, temperature, and physical location) and customer/third party data (e.g., CMDB or CMS as a service, Whois, geocoordinates, etc.). As discussed, the software sensors 112 and hardware sensors 114 can collect the captured data from multiple perspectives to provide a comprehensive view of network behavior. The software sensors 112 may include sensors along multiple points of a data path (e.g., network devices, physical or bare metals servers) and within multiple partitions of a physical host (e.g., hypervisor, container orchestrator, virtual entity manager, VM, container, other virtual entity, etc.).

After collection of the network telemetry and associated data, the method 600 may continue on to step 604, in which the application and network analytics platform can generate an application dependency map (ADM) for one or more applications executing in the network. In some embodiments, the nodes of the ADM can represent a pairing or concatenation of a server and a process (or application tier, application, application component, or other application granularity in other embodiments), and the edges of the ADM graph can represent the application and network analytics framework detecting flows between nodes. Clusters of nodes (i.e., host-process pairings within a similarity threshold of one another) can represent application components and each connected graph of the ADM can represent an application.

In some embodiments, generating the ADM can involve determining the nodes of the ADM by concatenating servers (e.g., the telemetry identifies a server as a source or destination in packet header metadata of a packet/flow) and applications/application components (e.g., the telemetry identifies a process generating a packet/flow). Generating the ADM can also include determining edges of the ADM by detecting one or more flows between nodes of the ADM. Generating the ADM can further include determining feature vectors for the nodes. The feature vector for each node may include a concatenation of server features, process features, and/or other features. The server features can include the server name, network address, operating system, CPU usage, network usage, disk space, ports, logged users, scheduled jobs, open files, and information regarding files and/or directories stored on the server. The process features can include the process name, process identifier, parent process identifier, process path, CPU utilization, memory utilization, memory address, scheduling information, nice value, flags, priority, status, start time, terminal type, CPU time taken by the process, the command string that started the process, and the process owner (e.g., user name, user identifier, user's real name, e-mail address, user's groups, terminal information, login time, expiration date of login, idle time, and information regarding files and/or directories of the user. In some embodiments, the feature vectors can also include features extracted from customer/third party data such as and customer/third party data (e.g., CMDB or CMS as a service, Whois, geocoordinates, etc.)

Telemetry used to detect flows between the nodes of the ADM may include packet header fields such as source address, source port, destination address, destination port, protocol type, class of service, etc. and/or aggregate packet data such as flow start time, flow end time, number of packets for a flow, number of bytes for a flow, the union of TCP flags for a flow, etc.

Generating the ADM can also involve determining similarity between the nodes of the ADM by comparing their feature vectors. Similarity can be a measure of how much alike two nodes are relative to other nodes, or a measure of two nodes being less distant to one another than other nodes. In some embodiments, the application and network analytics platform can use as similarity/distance measures one or more of Euclidean distance, Manhattan distance, Minkowski distance, cosine similarity, Jaccard similarity, and the like. In some embodiments, determining the ADM can further include setting the similarity threshold for clusters to specify a level of granularity with respect to a view of the applications executing in the network. For example, setting the similarity threshold to a very coarse degree of similarity can result in a single cluster representing the data center as a monolithic application. On the other hand, setting the similarity threshold to a very fine degree of similarity can result in singleton clusters for each host-process pairing in the network. Setting the similarity threshold can depend largely on the number and types of applications executing in the network and the level of granularity desired for the task at hand. In most situations, the similarity threshold may lie somewhere between very coarse and very fine. As discussed, clusters can represent application components, and a connected graph can represent an application.

After generating the ADM, the method 600 can proceed to step 606 in which the application and network analytics platform may devolve flows into one or more flowlets. That is, the application and network analytics platform can break a flow down into a series of sub-requests and sub-responses by tracing a flow from source to destination. A request flow can include hops over network devices from source to destination and processing of the flow by the network devices. A response flow can include hops over network devices, processing of the flow by the network devices, and sub-requests and sub-responses to intermediate endpoints (including hops over network devices and processing by these network devices) performed to generate a response to the originating request flow. For example, in FIG. 5, the application and network analytics platform may decompose the response of the web server 412 to the originating request from the client as a sequence including the segments BC, CD, DE, EF, FG, GH, HI, IJ, JK, KL, LM, MN, NO, OP, PQ, QR, ST, TU, UV, WX, XY, YZ, Zα, αβ, βγ, γδ, δε, and εζ.

The method 600 may continue to step 608 in which the application and network analytics platform can determine baseline metrics for the flowlets such as an average or a range of the length of time for transmitting and/or processing each flowlet and/or an average amount of data or a range of the amount of data (e.g., in number of packets and/or bytes) transmitted and/or processed within each flowlet. For example, the application and network analytics platform can detect a request flow associated with an application executing on a server using network telemetry and associated data captured by a software sensor of the server, such as by taking a snapshot of the processes executing on the server around or at the time that the server generated the flow and correlating the flow with one of the processes. The application and network analytics platform can trace the request flow from the server to one or more network devices (i.e., the network devices may correspond to one or more hops of the flow) and determine the transmission time(s) (and sometimes the amounts of data transmitted) between the server and the network device(s) as well as the time spent by the network device(s) to process the flow based on network telemetry captured by the hardware sensor(s) of the network device(s). The application and network analytics platform can trace the response flow as the transmission and/or processing of sub-requests and sub-responses to other servers (and transmission and/or processing by network devices along the data paths of the sub-requests and responses) based on network telemetry and associated data captured by the sensors of these additional devices and determine the amount of times (and/or the amount of data in some embodiments) associated with transmitting and/or processing the sub-requests and sub-responses. The application and network analytics platform can determine a baseline metric for a flowlet by aggregating observations of the flowlet detected during a specified period of time.

At step 610, the application and network analytics platform can evaluate network telemetry and associated data over a second period of time by comparing new flowlets against their corresponding baselines. In some embodiments, the application and network analytics platform may use machine learning and pattern recognition algorithms to classify each new flowlet. Machine learning can involve receiving “trained” samples (i.e., flowlets associated with predetermined classifications), determining the features (sometimes weighted) that the samples of a particular classification may have in common with one another and may not have in common with other classifications, and using these (sometimes weighted) features to classify new flowlets. Machine-learning and pattern recognition algorithms can include decision trees, random forests, rules-based classifiers, support vector machine (SVM) classifiers, neural network classifiers, nearest neighbor classifiers, etc.

Decision trees use a hierarchical division of the underlying data and different features. The hierarchical division of the data creates class partitions skewed in terms of their class distribution. For a given set of differences between flowlets, the partition that a flowlet is most likely to belong to is that flowlet's classification.

Random forests rely on a combination of decision trees in which each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. Implementers can train a random forest for a number of trees T by sampling n cases of the training data at random with replacement to create a subset of the training data. At each node, the classifier can select a number m of the features at random from the set of all features, and can ultimately select the feature that provides the best split to do a binary split on that node. At the next node, the classifier selects another number m of the features at random and repeats the process.

Rules-based classifiers determine patterns (e.g., regular expressions) that define the different classes. These types of classifiers define a set of rules in which the left-hand side corresponds to a pattern, and the right-hand side corresponds to a class label. The rules determine the classification.

SVM classifiers attempt to partition the underlying data using linear or non-linear delineations between the different classes. The premise of these classifiers is to determine the optimal boundaries between the different classes and use these boundaries for the purposes of classification.

Neural network classifiers utilize an interconnected group of functions or classifiers that process information using a connectionist approach. Neural networks change their structure during training, such as by merging overlapping detections within one network and training an arbitration network to combine the results from different networks.

Nearest neighbor classifiers determine the k-nearest neighbors in the training data to a specific set of differences between flowlets, and the majority class from the k neighbors determines the classification.

At decision point 612, if the new flowlets do not deviate from their baselines, such as by occurring within a specified time range and/or transmitting/processing an amount of data within a specified range of data sizes, then the application and network analytics platform may continue analyzing new flowlets.

However, if the application and network analytics platform determines that the new flowlets deviate from their baselines, at step 614, the application and network analytics platform can perform certain remediation tasks to address the deviations. In some embodiments, remediation can include generating an alert, such as via an event-based notification system (e.g., the event-based notification system 146 of FIG. 1). In some cases, a network administrator may analyze the anomalous flowlet and determine that performance of the application and network is in line with expectation or determine that the flowlet is indicative of an issue with application and network performance. The application and network analytics platform can receive this as input (i.e., a trained sample) to improve machine learning.

As discussed, in some embodiments, the application and network analytics platform can determine the physical and/or logical locations of servers within the network (e.g., CMDB/CMS, geocoordinates, IP address, etc.). The application and network analytics can use the location information to determine optimal placement and configuration of application components that are dependent on one another such that the components are closer together physically and/or logically. For example, the application and network analytics platform may determine that an anomalous flow is a bottleneck for a particular application and can locate a different location for the application component. As another example, the application and network analytics platform may have a comprehensive view of the application and network activity in the data center and can determine locations for application components that minimize bandwidth and latency. Thus, in some embodiments, remediation can include automatically migrating virtual entities having application dependencies closer together physically and/or logically to resolve application and network performance issues. In some embodiments, remediation can also include generating recommendations on physical placement or configuration of application components to be physically and/or logically closer together to reduce latency and other performance degradation.

FIG. 7A and FIG. 7B illustrate systems in accordance with various embodiments. The more appropriate system will be apparent to those of ordinary skill in the art when practicing the various embodiments. Persons of ordinary skill in the art will also readily appreciate that other systems are possible.

FIG. 7A illustrates an example architecture for a conventional bus computing system 700 wherein the components of the system are in electrical communication with each other using a bus 705. The computing system 700 can include a processing unit (CPU or processor) 710 and a system bus 705 that may couple various system components including the system memory 715, such as read only memory (ROM) in a storage device 770 and random access memory (RAM) 775, to the processor 710. The computing system 700 can include a cache 712 of high-speed memory connected directly with, in close proximity to, or integrated as part of the processor 710. The computing system 700 can copy data from the memory 715 and/or the storage device 730 to the cache 712 for quick access by the processor 710. In this way, the cache 712 can provide a performance boost that avoids processor delays while waiting for data. These and other modules can control the processor 710 to perform various actions. Other system memory 715 may be available for use as well. The memory 715 can include multiple different types of memory with different performance characteristics. The processor 710 can include any general purpose processor and a hardware module or software module, such as module 1 732, module 2 734, and module 3 736 stored in storage device 730, configured to control the processor 710 as well as a special-purpose processor where software instructions are incorporated into the actual processor design. The processor 710 may essentially be a completely self-contained computing system, containing multiple cores or processors, a bus, memory controller, cache, etc. A multi-core processor may be symmetric or asymmetric.

To enable user interaction with the computing system 700, an input device 745 can represent any number of input mechanisms, such as a microphone for speech, a touch-protected screen for gesture or graphical input, keyboard, mouse, motion input, speech and so forth. An output device 735 can also be one or more of a number of output mechanisms known to those of skill in the art. In some instances, multimodal systems can enable a user to provide multiple types of input to communicate with the computing system 700. The communications interface 740 can govern and manage the user input and system output. There may be no restriction on operating on any particular hardware arrangement and various other embodiments may substitute the basic features here for improved hardware or firmware arrangements.

Storage device 730 can be a non-volatile memory and can be a hard disk or other types of computer readable media which can store data that are accessible by a computer, such as magnetic cassettes, flash memory cards, solid state memory devices, digital versatile disks, cartridges, random access memories (RAMs) 725, read only memory (ROM) 720, and hybrids thereof.

The storage device 730 can include software modules 732, 734, 736 for controlling the processor 710. Various embodiments may utilize other hardware or software modules. The storage device 730 can connect to the system bus 705. In one aspect, a hardware module that performs a particular function can include the software component stored in a computer-readable medium in connection with the necessary hardware components, such as the processor 710, bus 705, output device 735, and so forth, to carry out the function.

FIG. 7B illustrates an example architecture for a conventional chipset computing system 750 that can be used in accordance with an embodiment. The computing system 750 can include a processor 755, representative of any number of physically and/or logically distinct resources capable of executing software, firmware, and hardware configured to perform identified computations. The processor 755 can communicate with a chipset 760 that can control input to and output from the processor 755. In this example, the chipset 760 can output information to an output device 765, such as a display, and can read and write information to storage device 770, which can include magnetic media, and solid state media, for example. The chipset 760 can also read data from and write data to RAM 775. A bridge 780 for connecting a variety of user interface components 785 can interface with the chipset 760. The user interface components 785 can include a keyboard, a microphone, touch detection and processing circuitry, a pointing device, such as a mouse, and so on. Inputs to the computing system 750 can come from any of a variety of sources, machine generated and/or human generated.

The chipset 760 can also interface with one or more communication interfaces 790 that can have different physical interfaces. The communication interfaces 790 can include interfaces for wired and wireless LANs, for broadband wireless networks, as well as personal area networks. Some applications of the methods for generating, displaying, and using the GUI disclosed herein can include receiving ordered datasets over the physical interface or be generated by the machine itself by processor 755 analyzing data stored in the storage device 770 or the RAM 775. Further, the computing system 700 can receive inputs from a user via the user interface components 785 and execute appropriate functions, such as browsing functions by interpreting these inputs using the processor 755.

It will be appreciated that computing systems 700 and 750 can have more than one processor 710 and 755, respectively, or be part of a group or cluster of computing devices networked together to provide greater processing capability.

For clarity of explanation, in some instances the various embodiments may be presented as including individual functional blocks including functional blocks comprising devices, device components, steps or routines in a method embodied in software, or combinations of hardware and software.

In some embodiments the computer-readable storage devices, mediums, and memories can include a cable or wireless signal containing a bit stream and the like. However, when mentioned, non-transitory computer-readable storage media expressly exclude media such as energy, carrier signals, electromagnetic waves, and signals per se.

Methods according to the above-described examples can be implemented using computer-executable instructions that are stored or otherwise available from computer readable media. Such instructions can comprise, for example, instructions and data which cause or otherwise configure a general purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Portions of computer resources used can be accessible over a network. The computer executable instructions may be, for example, binaries, intermediate format instructions such as assembly language, firmware, or source code. Examples of computer-readable media that may be used to store instructions, information used, and/or information created during methods according to described examples include magnetic or optical disks, flash memory, USB devices provided with non-volatile memory, networked storage devices, and so on.

Devices implementing methods according to these disclosures can comprise hardware, firmware, and/or software, and can take any of a variety of form factors. Typical examples of such form factors include laptops, smart phones, small form factor personal computers, personal digital assistants, rack mount devices, standalone devices, and so on. Other embodiments may implement the functionality described in this disclosure in peripherals or add-in cards. Various embodiments may also implement this functionality on a circuit board among different chips or different processes executing in a single device, by way of further example.

The instructions, media for conveying such instructions, computing resources for executing them, and other structures for supporting such computing resources are means for providing the functions described in these disclosures.

Although the disclosure uses a variety of examples and other information to explain aspects within the scope of the appended claims, a person having ordinary skill in art will understand not to limit the claims to particular features or arrangements in such examples, as one of ordinary skill can use these examples to derive a wide variety of implementations. Further, although the disclosure describes some subject matter in language specific to examples of structural features and/or method steps, one of ordinary skill will understand that the subject matter defined in the appended claims is not necessarily limited to these described features or acts. For example, such functionality can be distributed differently or performed in components other than those identified herein. Rather, the disclosure provides described features and steps as examples of components of systems and methods within the scope of the appended claims.

Claims

1. A method comprising: processing telemetry data for a plurality of flows associated with a set of service instances in a network, the set of service instances residing in a data center, the telemetry data received from one or more sensors installed in the data center;generating, based on the processed telemetry data, an application dependency map for an application executing in the network, the application dependency map indicating dependencies between the set of service instances in the network, each service instance implementing one or more processes associated with the application;determining one or more metrics associated with requests and responses transmitted between at least a first service instance and a second service instance of the application dependency map;comparing the determined one or more metrics to respective ranges; andresponsive to detecting a deviation of at least one of the determined one or more metrics from a corresponding respective range, initiating one or more remediation actions, at least one of the one or more remediation actions comprising instantiating one or more new service instances associated with the application in a public cloud remote from the data center.

2. The method of claim 1, wherein at least a second one of the one or more remediation actions comprises load balancing among the set of service instances associated with the application dependency map.

3. The method of claim 1, wherein at least one of the determined one or more metrics comprises central processing unit (CPU) utilization of the first service instance or the second service instance.

4. The method of claim 3, wherein at least a second one of the one or more remediation actions comprises instantiating one or more new service instances associated with the application dependency map in the data center.

5. The method of claim 1, wherein at least a second one of the one or more remediation actions comprises disabling network connectivity for one or more problematic servers.

6. The method of claim 1, wherein at least a second one of the one or more remediation actions comprises migrating the first service instance from a first location to a second location, wherein the migrating reduces a distance between the first service instance and a third location of the second service instance.

7. The method of claim 1, wherein at least one of the respective ranges is based on one or more baseline metrics, the one or more baseline metrics determined from analysis of the telemetry data over a period of time associated with the requests and responses transmitted between at least the first service instance and the second service instance of the application dependency map.

8. The method of claim 1, wherein at least one sensor of the one or more sensors is installed on a network device in the network, and wherein at least a second sensor of the one or more sensors is installed on a server device of the network.

9. A system comprising: one or more processors; andmemory including instructions that, upon being executed by the one or more processors, cause the system to: process telemetry data for a plurality of flows associated with a set of service instances in a network, the set of service instances residing in a data center, the telemetry data received from one or more sensors installed in the data center;generate, based on the processed telemetry data, an application dependency map for an application executing in the network, the application dependency map indicating dependencies between the set of service instances in the network, each service instance implementing one or more processes associated with the application;determine one or more metrics associated with requests and responses transmitted between at least a first service instance and a second service instance of the application dependency map;compare the determined one or more metrics to respective ranges; and responsive to detecting a deviation of at least one of the determined one or more metrics from a corresponding respective range, initiate one or more remediation actions, at least one of the one or more remediation actions comprising instantiating one or more new service instances associated with the application in a public cloud remote from the data center.

10. The system of claim 9, wherein at least a second one of the one or more remediation actions comprises load balancing among the set of service instances associated with the application dependency map.

11. The system of claim 9, wherein at least one of the determined one or more metrics comprises central processing unit (CPU) utilization of the first service instance or the second service instance.

12. The system of claim 11, wherein at least a second one of the one or more remediation actions comprises instantiating one or more new service instances associated with the application dependency map in the data center.

13. The system of claim 9, wherein at least a second one of the one or more remediation actions comprises disabling network connectivity for one or more problematic servers.

14. The system of claim 9, wherein at least a second one of the one or more remediation actions comprises migrating the first service instance from a first location to a second location, wherein the migrating reduces a distance between the first service instance and a third location of the second service instance.

15. The system of claim 9, wherein at least one of the respective ranges is based on one or more baseline metrics, the one or more baseline metrics determined from analysis of the telemetry data over a period of time associated with the requests and responses transmitted between at least the first service instance and the second service instance of the application dependency map.

16. The system of claim 9, wherein at least one sensor of the one or more sensors is installed on a network device in the network, and wherein at least a second sensor of the one or more sensors is installed on a server device of the network.

17. A non-transitory computer-readable medium having instructions that, upon being executed by one or more processors, cause the one or more processors to: process telemetry data for a plurality of flows associated with a set of service instances in a network, the set of service instances residing in a data center, the telemetry data received from one or more sensors installed in the data center;generate, based on the processed telemetry data, an application dependency map for an application executing in the network, the application dependency map indicating dependencies between the set of service instances in the network, each service instance implementing one or more processes associated with the application;determine one or more metrics associated with requests and responses transmitted between at least a first service instance and a second service instance of the application dependency map;compare the determined one or more metrics to respective ranges; andresponsive to detecting a deviation of at least one of the determined one or more metrics from a corresponding respective range, initiate one or more remediation actions, at least one of the one or more remediation actions comprising instantiating one or more new service instances associated with the application in a public cloud remote from the data center.

18. The non-transitory computer-readable medium of claim 17, wherein at least a second one of the one or more remediation actions comprises load balancing among the set of service instances associated with the application dependency map.

19. The non-transitory computer-readable medium of claim 17, wherein at least one of the determined one or more metrics comprises central processing unit (CPU) utilization of the first service instance or the second service instance.

20. The non-transitory computer-readable medium of claim 19, wherein at least a second one of the one or more remediation actions comprises instantiating one or more new service instances associated with the application dependency map in the data center.

21. The non-transitory computer-readable medium of claim 17, wherein at least a second one of the one or more remediation actions comprises disabling network connectivity for one or more problematic servers.

22. The non-transitory computer-readable medium of claim 17, wherein at least a second one of the one or more remediation actions comprises migrating the first service instance from a first location to a second location, wherein the migrating reduces a distance between the first service instance and a third location of the second service instance.