Skip to content

Async Mode Nginx with QAT support which improves Crypto and compression performance

License

Notifications You must be signed in to change notification settings

intel/asynch_mode_nginx

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Asynch Mode for NGINX*

Nginx ("engine x") is an HTTP web server, reverse proxy, content cache, load balancer, TCP/UDP proxy server, and mail proxy server. Originally written by Igor Sysoev and distributed under the 2-clause BSD License. This project provides async capabilities to Nginx using OpenSSL* ASYNC infrastructure. The Async mode Nginx* with Intel® QuickAssist Technology (QAT) Acceleration can provide significant performance improvements.

Table of Contents

Features

  • Asynchronous Mode in SSL/TLS processing (including http/stream/mail/proxy module)
  • Support QAT_HW and QAT_SW PKE and symmetric Crypto acceleration
  • Support QAT_HW GZIP compression acceleration
  • SSL Engine Framework for engine configuration
  • Support for external polling mode and heuristic polling mode
  • Release hardware resource during worker is shutting down (For more details information, please read modules/nginx_qat_module/README)
  • Support OpenSSL Cipher PIPELINE Capability.
  • Support software fallback for asymmetric cryptography algorithms(Linux & FreeBSD) and symmetric algorithms (FreeBSD only)

Hardware Requirements

Asynch Mode for NGINX* supports QAT_HW Crypto and Compression acceleration on the platforms with the following QAT devices

QAT_SW acceleration is supported in the platforms starting with 3rd Generation Intel® Xeon® Scalable Processors family and later.

Software Requirements

This release was validated on the following and its dependent libraries mentioned QAT Engine and QATZip software Requirements section.

Installation Instructions

Pre-requisties: Install OpenSSL and QAT Engine(for Crypto using QAT_HW & QAT_SW) and QATZip(for Compression) using the QAT engine and QATzip installation instructions.

Set the following environmental variables:

NGINX_INSTALL_DIR is the directory where Nginx will be installed to

OPENSSL_LIB is the directory where the OpenSSL has been installed to

QZ_ROOT is the directory where the QATzip has been compiled to

Configure and build nginx with QAT Engine and QATZip

./configure \
  --prefix=$NGINX_INSTALL_DIR \
  --with-http_ssl_module \
  --add-dynamic-module=modules/nginx_qatzip_module \
  --add-dynamic-module=modules/nginx_qat_module/ \
  --with-cc-opt="-DNGX_SECURE_MEM -I$OPENSSL_LIB/include -I$ICP_ROOT/quickassist/include -I$ICP_ROOT/quickassist/include/dc -I$QZ_ROOT/include -Wno-error=deprecated-declarations" \
  --with-ld-opt="-Wl,-rpath=$OPENSSL_LIB/lib64 -L$OPENSSL_LIB/lib64 -L$QZ_ROOT/src -lqatzip -lz"
  make
  make install
  • Refer QAT Settings here for running Nginx under non-root user.

Testing

Official Unit tests

These instructions can be found on Official test

Performance Testing

Performance Testing with Async mode Nginx along with best known build and runtime configuration along with scripts available here

QAT Configuration

Refer Sample Configuration file conf/nginx.QAT-sample.conf with QAT configured which includes

  • SSL QAT engine in heuristic mode.

  • HTTPS in asynchronous mode.

  • QATZip Module for compression acceleration.

  • TLS1.3 Support as default.

  • The QAT Hardware driver configuration needs crypto and compression instance configured

SSL Engine (QAT Engine) configuration

An SSL Engine Framework is introduced to provide a more powerful and flexible mechanism to configure Nginx SSL engine directly in the Nginx configuration file (nginx.conf).

Loading Engine via OpenSSL.cnf is not viable in Nginx. Please use the SSL Engine Framework which provides a more powerful and flexible mechanism to configure Nginx SSL engine such as Intel® QAT Engine directly in the Nginx configuration file (nginx.conf).

A new configuration block is introduced as ssl_engine which provides two general directives:

Sets the engine module and engine id for OpenSSL async engine. For example:

Syntax:     use_engine [engine_module_name] [engine_id];
Default:    N/A
Context:    ssl_engine
Description:
            Specify the engine module name against engine id

where engine_module_name is optional and should be same as engine id if it is not used.

Syntax:     default_algorithms [ALL/RSA/EC/...];
Default:    ALL
Context:    ssl_engine
Description:
            Specify the algorithms need to be offloaded to the engine
            More details information please refer to OpenSSL engine

If Handshake only acceleration is preferred then default_algorithms should RSA EC DH DSA PKEY and if symmetric only is preferred then use CIPHERS. If all algorithms are preferred then ALL must be specified.

A following configuration sub-block can be used to set engine specific configuration. The name of the sub-block should have a prefix using the engine name specified in use_engine, such as [engine_name]_engine.

Any 3rd party modules can be integrated into this framework. By default, a reference module dasync_module is provided in src/engine/modules and a QAT module nginx_qat_module is provided in modules/nginx_qat_modules.

An example configuration in the nginx.conf:

    load_module modules/ngx_ssl_engine_qat_module.so;
    ...
    ssl_engine {
        use_engine qatengine;
        default_algorithms RSA,EC,DH,PKEY_CRYPTO;
        qat_engine {
            qat_sw_fallback on;
            qat_offload_mode async;
            qat_notify_mode poll;
            qat_poll_mode heuristic;
            qat_shutting_down_release on;
        }
    }

For more details directives of nginx_qat_module, please refer to modules/nginx_qat_modules/README.

Nginx Side Polling

The qat_module provides two kinds of Nginx side polling for QAT engine,

  • external polling mode
  • heuristic polling mode

Please refer to the README file in the modules/nginx_qat_modules directory to install this dynamic module.

Note: External polling and heuristic polling are unavailable in SSL proxy and stream module up to current release.

External Polling Mode

A timer-based polling is employed in each Nginx worker process to collect accelerator's response.

Directives in the qat_module

Syntax:     qat_external_poll_interval time;
Default:    1
Dependency: Valid if (qat_poll_mode=external)
Context:    qat_engine
Description:
            External polling time interval (ms)
            Valid value: 1 ~ 1000

Example file: conf/nginx.conf

    load_module modules/ngx_ssl_engine_qat_module.so;
    ...
    ssl_engine {
        use_engine qatengine;
        default_algorithms ALL;
        qat_engine {
            qat_offload_mode async;
            qat_notify_mode poll;

            qat_poll_mode external;
            qat_external_poll_interval 1;
        }
    }

Heuristic Polling Mode

This mode can be regarded as an improvement of the timer-based external polling. It is the recommended polling mode to communicate with QAT accelerators because of its performance advantages. With the knowledge of the offload traffic, it controls the response reaping rate to match the request submission rate so as to maximize system efficiency with moderate latency, and adapt well to diverse types of network traffics.

Note:

  • This mode is only available when using QAT engine v0.5.35 or later.
  • External polling timer is enabled by default when heuristic polling mode is enabled.

In the heuristic polling mode, a polling operation is only triggered at a proper moment:

  • Number of in-flight offload requests reaches a pre-defined threshold (in consideration of efficiency)
  • All the active SSL connections have submitted their cryptographic requests and been waiting for the corresponding responses (in consideration of timeliness).

Directives in the qat_module

Syntax:     qat_heuristic_poll_asym_threshold num;
Default:    48
Dependency: Valid if (qat_poll_mode=heuristic)
Context:    qat_engine
Description:
            Threshold of the number of in-flight requests to trigger a polling
            operation when there are in-flight asymmetric crypto requests
            Valid value: 1 ~ 512


Syntax:     qat_heuristic_poll_sym_threshold num;
Default:    24
Dependency: Valid if (qat_poll_mode=heuristic)
Context:    qat_engine
Description:
            Threshold of the number of in-flight requests to trigger a polling
            operation when there is no in-flight asymmetric crypto request
            Valid value: 1 ~ 512

Example file: conf/nginx.conf

    load_module modules/ngx_ssl_engine_qat_module.so;
    ...
    ssl_engine {
        use_engine qatengine;
        default_algorithms ALL;
        qat_engine {
            qat_offload_mode async;
            qat_notify_mode poll;

            qat_poll_mode heuristic;
            qat_heuristic_poll_asym_threshold 48;
            qat_heuristic_poll_sym_threshold 24;
        }
    }

QATzip Module Configuration

This module is developed to accelerate GZIP compression with QAT accelerators through QATzip stream API released in v0.2.6.

Note:

  • This mode is only available when using QATzip v1.0.0 or later.
  • This mode relies on gzip module for SW fallback feature.
  • The qatzip_sw is set to failover by default, do not load QATzip module if you do not want to enable qatzip. Or else it would be enabled and set to failover.

Directives in the qatzip_module

    Syntax:     qatzip_sw only/failover/no;
    Default:    qatzip_sw failover;
    Context:    http, server, location, if in location
    Description:
                only: qatzip is disable, using gzip;
                failover: qatzip is enable, qatzip sfotware fallback feature enable.
                no: qatzip is enable, qatzip sfotware fallback feature disable.

    Syntax:     qatzip_chunk_size size;
    Default:    qatzip_chunk_size 64k;
    Context:    http, server, location
    Description:
                Sets the chunk buffer size in which data will be compressed into
                one deflate block. By default, the buffer size is equal to 64K.

    Syntax:     qatzip_stream_size size;
    Default:    qatzip_stream_size 256k;
    Context:    http, server, location
    Description:
                Sets the size of stream buffers in which data will be compressed into
                multiple deflate blocks and only the last block has FINAL bit being set.
                By default, the stream buffer size is 256K.

Example file: conf/nginx.conf

    load_module modules/ngx_http_qatzip_filter_module.so;
    ...

    gzip_http_version   1.0;
    gzip_proxied any;
    qatzip_sw failover;
    qatzip_min_length 128;
    qatzip_comp_level 1;
    qatzip_buffers 16 8k;
    qatzip_types text/css text/javascript text/xml text/plain text/x-component application/javascript application/json application/xml application/rss+xml font/truetype font/opentype application/vnd.ms-fontobject image/svg+xml application/octet-stream image/jpeg;
    qatzip_chunk_size   64k;
    qatzip_stream_size  256k;
    qatzip_sw_threshold 256;

For more details directives of nginx_qatzip_module, please refer to modules/nginx_qatzip_modules/README.

Additional Information

Generating Async mode Patch

  • Asynch Mode for NGINX* is developed based on Nginx-1.26.2. Refer steps below to generate patch to apply on top of official Nginx-1.26.2.

Generating the patch with package from Nginx download page

  git clone https://github.com/intel/asynch_mode_nginx.git asynch_mode_nginx
  wget http://nginx.org/download/nginx-1.26.2.tar.gz
  tar -xvzf nginx-1.26.2.tar.gz
  diff -Naru -x .git nginx-1.26.2 asynch_mode_nginx > async_mode_nginx_1.26.2.patch
  # Apply Patch
  cd nginx-1.26.2
  patch -p1 < ../async_mode_nginx_1.26.2.patch

Generating the patch with package from Github Mirror page

  git clone https://github.com/intel/asynch_mode_nginx.git
  wget https://github.com/nginx/nginx/archive/release-1.26.2.tar.gz
  tar -xvzf release-1.26.2.tar.gz
  diff -Naru -x .git -x .hgtags nginx-release-1.26.2 asynch_mode_nginx > async_mode_nginx_1.26.2.patch
  # Apply Patch
  patch -p0 < async_mode_nginx_1.26.2.patch

New Async Directives

Directives

Syntax:     ssl_asynch on | off;
Default:    ssl_asynch off;
Context:    stream, mail, http, server

    Enables SSL/TLS asynchronous mode

Example

file: conf/nginx.conf

    http {
        ssl_asynch  on;
        server {...}
    }

    stream {
        ssl_asynch  on;
        server {...}
    }

Directives

Syntax:     proxy_ssl_asynch on | off;
Default:    proxy_ssl_asynch off;
Context:    stream, http, server

Enables the SSL/TLS protocol asynchronous mode for connections to a proxied server.

Example

file: conf/nginx.conf

    http {
        server {
            location /ssl {
                proxy_pass https://127.0.0.1:8082/outer;
                proxy_ssl_asynch on;
            }
        }
    }

Directives

Syntax:     grpc_ssl_asynch on | off;
Default:    grpc_ssl_asynch off;
Context:    http, server

Enables the SSL/TLS protocol asynchronous mode for connections to a grpc server.

Example

file: conf/nginx.conf

    http {
        server {
            location /grpcs {
                grpc_pass https://127.0.0.1:8082/outer;
                grpc_ssl_asynch on;
            }
        }
    }
  • OpenSSL Cipher PIPELINE Capability. Refer OpenSSL Docs for detailed information.

Directives

Syntax:     ssl_max_pipelines size;
Default:    ssl_max_pipelines 0;
Context:    server

Set MAX number of pipelines

Syntax:     ssl_split_send_fragment size;
Default:    ssl_split_send_fragment 0;
Context:    server

Set split size of sending fragment

Syntax:     ssl_max_send_fragment size;
Default:    ssl_max_send_fragment 0;
Context:    server

Set max number of sending fragment

Limitations

  • Nginx supports reload operation, when QAT hardware is involved for crypto offload, user should ensure that there are enough number of QAT instances. For example, the available qat instance number should be 2x equal or more than Nginx worker process number.

    For example, in Nginx configuration file (nginx.conf) worker process number is configured as

       worker_processes 16;

    Then the instance configuration in QAT driver configuration file should be

        [SHIM]
        NumberCyInstances = 1
        NumberDcInstances = 0
        NumProcesses = 32
        LimitDevAccess = 0
  • When configure "worker_process auto", Asynch Mode for NGINX* will need instance number equal or larger than 2 times of CPU core number. Otherwise, Asynch Mode for NGINX* might show various issue caused by leak of instance.

  • Nginx supports QAT engine and QATzip module. By default, they use User Space DMA-able Memory (USDM) Component. The USDM component is located within the Upstream Intel® QAT Driver source code in the following subdirectory: quickassist/utilities/libusdm_drv. We should have this component enabled and assigned enough memory before using nginx_qat_module or nginx_qatzip_module, for example:

        echo 2048 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages
        insmod ./usdm_drv.ko max_huge_pages=2048 max_huge_pages_per_process=32
  • AES-CBC-HMAC-SHA algorithm won't be offloaded to QAT_HW if Encrypt then MAC(ETM) mode is used for SSL connection(by default). Refer to QAT_Engine Limitations for more details.

  • QATzip module supports GZIP compression acceleration now, does not support user define dictionary compression yet.

  • The HTTP2 protocol does not support asynchronous functionality.

  • QUIC is not supported for Async nginx with QAT.

Known Issues

'Orphan ring' errors in dmesg output when Nginx exit
Working with current QAT driver, Nginx workers exit with 'Orphan ring' errors. This issue has been fixed in future QAT driver release.

Cache manager/loader process will allocate QAT instance via QAT engine
According to current QAT engine design, child process forked by master process will initialize QAT engine automatically in QAT engine atfork hook function. If cache manager/loader processes are employed, QAT instances will be allocated in the same way as worker process. Cache manager/loader processes do not perform modules' exit process method in Nginx native design which will introduce "Orphan ring" error message in dmesg output.

Segment fault happens while sending HUP signal when QAT instances not enough
If the available qat instance number is less than 2x Nginx worker process number, segment fault happens while sending HUP signal to Asynch Mode for NGINX*. Using qat_sw_fallback on; in qat_engine directive as a workaround for this issue. And it needs special attention if the QAT instances are enough when setting worker_processes auto;.

Insufficient HW resources would cause segment fault while sending HUP signal
Before running nginx, please make sure you have arranged enough HW resources for nginx, including memory and hard disk space. Disk space exhausted or out of memory would cause core dump when nginx receives HUP signal during handshake phase.

Performance drop under OpenSSL 3.0

  • Both ECDH and PRF cause performance drop under OpenSSL 3.0.
  • Due to changes in the OpenSSL 3.0 framework, TLS handshake performance is significantly degraded. Refer #21833 and associated issues for more details.

The 0-RTT (early data) issue
The 0-RTT (early data) feature does not support async mode in current asynch_mode_nginx, so it's not recommended to use async offload to QAT hardware during early data process.

CHACHA-POLY and AES-GCM throughput performance issue
With the bottleneck of memory allocation, the throughput of CHACHA-Poly and AES-GCM can not reach the peak value when running with 4 or more QAT devices.

Licensing

Released under BSD License. Please see the LICENSE file contained in the top level folder. Further details can be found in the file headers of the relevant files.

Legal

Intel, Intel Atom, and Xeon are trademarks of Intel Corporation in the U.S. and/or other countries.

*Other names and brands may be claimed as the property of others.

Copyright © 2014-2024, Intel Corporation. All rights reserved.