Tag Archives: performance

Scaling out SSL

Synopsis

We’ve seen recently how we could scale up SSL performance.
But what about scaling out SSL performance?
Well, thanks to Aloha and HAProxy, it’s easy to manage smartly a farm of SSL accelerator servers, using persistence based on the SSL Session ID.
This way of load-balancing is smart, but in case of SSL accelerator failure, other servers in the farm would have a CPU overhead to generate SSL Session IDs for sessions re-balanced by the Aloha.

After a talk with (the famous) emericbr, HAProxy Technologies dev team leader, he decided to write a patch for stud to add a new feature: sharing SSL session between different stud processes.
That way, in case of SSL accelerator failure, the servers getting re-balanced sessions would not have to generate a new SSL session.

Emericbr’s patch is available here: https://github.com/bumptech/stud/pull/50
At the end of this article, you’ll learn how to use it.

Stud SSL Session shared caching

Description

As we’ve seen in our article on SSL performance, a good way to improve performance on SSL is to use a SSL Session ID cache.

The idea here, is to use this cache as well as sending updates into a shared cache one can consult to get the SSL Session ID and the data associated to it.

As a consequence, there are 2 levels of cache:

      * Level 1: local process cache, with the currently used SSL session
      * Level 2: shared cache, with the SSL session from all local cache

Way of working

The protocol understand 3 types of request:

      * New: When a process generates a new session, it updates its local cache then the shared cache
      * Get: When a client tries to resume a session and the process receiving it is not aware of it, then the process tries to get it from the shared cache
      * Del: When a session has expired or there is a bad SSL ID sent by a client, then the process will delete the session from the shared cache

Who does what?

Stud has a Father/Son architecture.
The Father starts up then starts up Sons. The Sons bind the TCP external ports, load the certificate and process the SSL requests.
Each son manages its local cache and send the updates to the shared cache. The Father manages the shared cache, receiving the changes and maintaining it up to date.

How are the updates exchanged?

Updates are sent either on Unicast or Multicast, on a specified UDP port.
Updates are compatible both IPv4 and IPv6.
Each packet are signed by an encrypted signature using the SSL certificate, to avoid cache poisoning.

What does a packet look like?


SSL Session ID ASN-1 of SSL Session structure Timestamp Signature
[32 bytes] [max 512 bytes] [4 bytes] [20 bytes]

Note: the SSL Session ID field is padded with 0 if required

Diagram

Let’s show this on a nice picture where each potato represents each process memory area.
stud_shared_cache
Here, the son on host 1 got a new SSL connection to process, since he could not find it in its cache and in the shared cache, he generated the asymmetric key, then push it to his shared cache and the father on host 2 which updates the shared cache for this host.
That way, if this user is routed to any stud son process, he would not have to compute again its asymmetric key.

Let’s try Stud shared cache

Installation:

git clone https://github.com/EmericBr/stud.git
cd stud
wget http://1wt.eu/tools/ebtree/ebtree-6.0.6.tar.gz
tar xvzf ebtree-6.0.6.tar.gz
ln -s ebtree-6.0.6 ebtree
make USE_SHARED_CACHE=1

Generate a key and a certificate, add them in a single file.

Now you can run stud:

sudo ./stud -n 2 -C 10000 -U 10.0.3.20,8888 -P 10.0.0.17 -f 10.0.3.20,443 -b 10.0.0.3,80 cert.pem

and run a test:

curl --noproxy * --insecure -D - https://10.0.3.20:443/

And you can watch the synchronization packets:

$ sudo tcpdump -n -i any port 8888
[sudo] password for bassmann: 
tcpdump: verbose output suppressed, use -v or -vv for full protocol decode
listening on any, link-type LINUX_SLL (Linux cooked), capture size 65535 bytes

17:47:10.557362 IP 10.0.3.20.8888 > 10.0.0.17.8888: UDP, length 176
17:49:04.592522 IP 10.0.3.20.8888 > 10.0.0.17.8888: UDP, length 176
17:49:05.476032 IP 10.0.3.20.8888 > 10.0.0.17.8888: UDP, length 176

Related links

Benchmarking SSL performance

Introduction

The story

Recently, there has been some attacks against website which aimed to steal user identity. In order to protect their users, major website owners had to find a solution.
Unfortunately, we know that sometimes, improving security means downgrading performance.

SSL/TLS is a fashion way to improve data safety when data is exchanged over a network.
SSL/TLS encryption is used to crypt any kind of data, from the login/password on a personnal blog service to a company extranet passing through an e-commerce caddy.
Recent attack shown that to be protect users identity, all the traffic must be encrypted.

Note, SSL/TLS is not only used on Website, but can be used to crypt any TCP based protocol like POP, IMAP, SMTP, etc…

Why this benchmark?

At HAProxy Technologies, we build load-balancer appliances based on a Linux kernel, LVS (for layer 3/4 load-balancing), HAProxy (for layer 7 load-balancing) and stunnel (SSL encryption), for the main components.

  1. Since SSL/TLS is fashion, we wanted to help people ask the right questions and to do the right choice when they have to bench and choose SSL/TLS products.
  2. We wanted to explain to everybody how one can improve SSL/TLS performance by adding some functionality to SSL open source software.
  3. Lately, on HAProxy mailing list, Sebastien introduced us to stud, a very good, but still young, alternative to stunnel. So we were curious to bench it.

SSL/TLS introduction

The theory

SSL/TLS can be a bit complicated at first sight.
Our purpose here is not to describe exactly how it works, there are useful readings for that:

    SSL main lines

    Basically, there are two main phases in SSL/TLS:

    1. the handshake
    2. data exchange

    During the handshake, the client and the server will generate three keys which are unique for the client and the server, available during the session life only and will be used to crypt and uncrypt data on both sides by the symmetric algorithms.
    Later in the article, we will use the term “Symmetric key” for those keys.

    The symmetric key is never exchanged over the network. An ID, called SSL session ID, is associated to it.

    Let’s have a look at the diagram below, which shows a basic HTTPS connection, step by step:

    SSL_handshake

    We also represented on the diagram the factor which might have an impact on performance.

    1. Client sends to the server the Client Hello packet with some randon numbers, its supported ciphers and a SSL session ID in case of resuming SSL session
    2. Server chooses a cipher from the client cipher list and sends a Server Hello packet, including random number.
      It generates a new SSL session ID if resume is not possible or available.
    3. Server sends its public certificate to the client, the client validates it against CA certificates.
      ==> sometimes you may have warnings about self-signed certificates.
    4. Server sends a Server Hello Done packet to tell the client he has finished for now
    5. Client generates and sends the pre-master key to the server
    6. Client and server generate the symmetric key that will be used to crypt data
    7. Client and server tells each other the next packets will be sent encrypted
    8. Now, data is encrypted.

    SSL Performance

    As you can see on the diagram above, some factors may influence SSL/TLS performance on the server side:

    1. the server hardware, mainly the CPU
    2. the asymmetric key size
    3. the symmetric algorithm

    In this article, we’re going to study the influence of these 4 factors and observe the impact on the performance.

    A few other things might have an impact on performance:

    • the ability to resume a SSL/TLS session
    • symmetric key generation frequency
    • object size to crypt

    Benchmark platform

    We used the platform below to run our benchmark:

    SSL_benchmark_platform

    The SSL server has purposely much less capacity than the client in order to ensure the client won’t saturate before the server.

    The client is inject + stunnel on client mode.
    The web server behind HAProxy and the SSL offloader is httpterm

    Note: Some resuts were checked using httperf and curl-loader, and results were similar.

    On the server, we have 2 cores and since we have enabled hyper threading, we have 4 CPUs available from a kernel point of view.
    The e1000e driver of the server has been modified to be able to bind interrupts on the first logical CPU core 0.

    Last but not least, the SSL library used is Openssl 0.9.8.

    Benchmark purpose

    The purpose of this benchmark is to:

    • Compare the different way of working of stunnel (fork, pthread, ucontext, ucontext + session cache)
    • Compare the different way of working of stud (without and with session cache)
    • Compare stud and stunnel (without and with session cache)
    • Impact of session renegotiation frequency
    • Impact of asymmetric key size
    • Impact of object size
    • Impact of symmetric cypher

    At the end of the document, we’re going to give some conclusion as well as some advices.

    As a standard test, we’re going to use the following:
    Protocol: TLSv1
    Asymmetric key size: 1024 bits
    Cipher: AES256-SHA
    Object size: 0 byte

    For each test, we’re going to provide the transaction per second (TPS) and the handshake capacity, which are the two most important numbers you need to know when comparing SSL accelerator products.

    • Transactions per second: the client will always re-use the same SSL session ID
    • Symmetric key generation: the client will never re-use its SSL session ID, forcing the server to generate a new symmetric key for each request

    1. From the influence of symmetric key generation frequency

    For this test, we’re going to use the following parameters:
    Protocol: TLSv1
    Asymmetric key size: 1024 bits
    Cipher: AES256-SHA
    Object size: 0 byte
    CPU: 1 core

    Note that the object is void because we want to mesure pure SSL performance.

    We’re going to bench the following software:
    STNL/FORK: stunnel-4.39 mode fork
    STNL/PTHD: stunnel-4.39 mode pthread
    STNL/UCTX: stunnel-4.39 mode ucontext
    STUD/BUMP: stud github bumptech (github: 1846569)

    Symmetric key generation frequency STNL/FORK STNL/PTHD STNL/UCTX STUD/BUMP
    For each request 131 188 190 261
    Every 100 requests 131 487 490 261
    Never 131 495 496 261

    Observation:

    – We can clearly see that STNL/FORK and STUD/BUMP can’t resume a SSL/TLS session.
    STUD/BUMP has better performance than STNL/* on symmetric key generation.

    2. From the advantage of caching SSL session

    For this test, we have developed patches for both stunnel and stud to improve a few things.
    The stunnel patches are applied on STNL/UCTX and include:
    – listen queue settable
    – performance regression due to logs fix
    – multiprocess start up management
    – session cache in shared memory

    The stud patches are applied on STUD/BUMP and include:
    – listen queue settable
    – session cache in shared memory
    – fix to allow session resume

    We’re going to use the following parameters:
    Protocol: TLSv1
    Asymmetric key size: 1024 bits
    Cipher: AES256-SHA
    Object size: 0 byte
    CPU: 1 core

    Note that the patched version will be respectively called STNL/PATC and STUD/PATC in the rest of this document.
    The percentage highlights the improvement of STNL/PATC and STUD/PATC respectively over STNL/UCTX and STUD/BUMP.

    Symmetric key generation frequency STNL/PATC STUD/PATC
    For each request 246
    +29%
    261
    +0%
    Every 100 requests 1085
    +121%
    1366
    +423%
    Never 1129
    +127%
    1400
    +436%

    Observation:

    – obviously, caching SSL session improves the number of transaction per second
    stunnel patches also improved stunnel performance

    3. From the influence of CPU cores

    As seen on the previous test, we could improve TLS capacity by adding a symmetric key cache to both stud and stunnel.
    We still might be able to improve things :).

    For this test, we’re going to configure both stunnel and stud to use 2 CPU cores.
    The kernel will be configured on core 0, userland on core 1 and stunnel or stud on cores 2 and 3, as shown below:
    cpu_affinity_ssl_2_cores

    For the rest of the tests, we’re going to bench only STNL/PTHD, which is the stunnel mode used by most of linux distribution, and the two patched STNL/PATC and STUD/PATC.

    For this test, we’re going to use the following parameters:
    Protocol: TLSv1
    Asymmetric key size: 1024 bits
    Cipher: AES256-SHA
    Object size: 0 byte
    CPU: 2 cores

    The table below summarizes the number we get with 2 cores and the percentage of improvement with 1 core:

    Symmetric key generation frequency STNL/PTHD STNL/PATC STUD/PATC
    For each request 217
    +15%
    492
    +100%
    517
    +98%
    Every 100 requests 588
    +20%
    2015
    +85%
    2590
    +89%
    Never 602
    +21%
    2118
    +87%
    2670
    +90%

    Observation:

    – now, we know the number of CPU cores has an influence 😉
    – the symmetric key generation has doubled on the patched versions. STNL/FORK does not take advantage of the second CPU core.
    – we can clearly see the benefit of SSL session caching on both STNL/PATC and STUD/PATC
    STUD/PATC performs around 25% better than STNL/PATC

    Note that since STNL/FORK and STUD/BUMP have no SSL session cache, no need to test them anymore.
    We’re going to concentrate on STNL/PTHD, STNL/UCTX, STNL/PATC and STUD/PATC.

    4. From the influence of the asymmetric key size

    The default asymmetric key size on current website is usually 1024 bits. For security purpose, more and more engineer now recommend using 2048 bits or even 4096 bits.
    In the following test, we’re going to use observe the impact of the asymmetric key size on the SSL performance.

    For this test, we’re going to use the following parameters:
    Protocol: TLSv1
    Asymmetric key size: 2048 bits
    Cipher: AES256-SHA
    Object size: 0 byte
    CPU: 2 cores

    The table below summarizes the number we got with 2048 bits asymmetric key size generation and the percentage highlights the performance impact compared to the 1024 bits asymmetric key size, both tests running on 2 CPU cores:

    Symmetric key generation frequency STNL/PTHD STNL/PATC STUD/PATC
    For each request 46
    -78%
    96
    -80%
    96
    -81%
    Every 100 requests 541
    -8%
    1762
    -13%
    2121
    -18%
    Never 602
    +0%
    2118
    0%
    2670
    +0%

    Observation:

    – the asymmetric key size has only an influence on symmetric key generation. The number of transaction per second does not change at all for the software which are able to cache and re-use SSL session id.
    – passing from 1024 to 2048 bits means dividing by 4 the number of symmetric key generated per second on our environment.
    – on an average traffic with renegotiation every 100 requests, stud is more impacted than stunnel but it performs better anyway.

    5. From the influence of the object size

    If you read carefully the article since the beginning, then you might be thinking “they’re nice with their test, but thier objects are empty… what happens with real objects?”
    So, I guess it’s time to study the impact of the object size!

    For this test, we’re going to use the following parameters:
    Protocol: TLSv1
    Asymmetric key size: 1024 bits
    Cipher: AES256-SHA
    Object size: 1 KByte / 4 KBytes
    CPU: 2 cores

    Results for STNL/PTHD, STNL/PATC and STUD/PATC:
    The percentage number highlights the performance impact.

    Symmetric key generation frequency STNL/PTHD STNL/PATC STUD/PATC
       1KB       4KB       1KB       4KB       1KB       4KB   
    every 100 requests 582 554
    -5%
    1897 1668
    -13%
    2475 2042
    -21%
    never 595 564
    -5%
    1997 1742
    -14%
    2520 2101
    -19%

    Observation

    – the bigger the object, the lower the performance…
    To be fair, we’re not surprised by this result 😉
    STUD/PATC performs 20% better than STNL/PATC
    STNL/PATC performs 3 times better than STNL/PTHD

    6. From the influence of the cipher

    Since the beginning, we run our bench only with the cipher AES256-SHA.
    It’s now the time to bench some other cipher:
    – first, let’s give a try to AES128-SHA, and compare it to AES256-SHA
    – second, let’s try RC4_128-SHA, and compare it to AES128-SHA

    For this test, we’re going to use the following parameters:
    Protocol: TLSv1
    Asymmetric key size: 1024 bits
    Cipher: AES256-SHA / AES128-SHA / RC4_128-SHA
    Object size: 4Kbyte
    CPU: 2 cores

    Results for STNL/PTHD, STNL/PATC and STUD/PATC:
    The percentage number highlights the performance impact on the following cipher:
    – AES 128 ==> AES 256
    – RC4 128 ==> AES 128

    Symmetric key generation frequency STNL/PTHD STNL/PATC STUD/PATC
    AES256 AES128 RC4_128 AES256 AES128 RC4_128 AES256 AES128 RC4_128
    every 100 requests 554 567
    +2%
    586
    +3%
    1668 1752
    +5%
    1891
    +8%
    2042 2132
    +4%
    2306
    +8%
    never 564 572
    +1%
    600
    +5%
    1742 1816
    +4%
    1971
    +8%
    2101 2272
    +8%
    2469
    +8%

    Observation:

    – As expected, AES128 performs better than AES256
    RC4 128 performs better than AES128
    stud performs better than stunnel
    – Note that RC4 will perform better on big objects, since it works on a stream while AES works on blocks

    Conclusion on SSL performance

    1.) bear in mind to ask the 2 numbers when comparing SSL products:
    – the number of handshakes per second
    – the number of transaction per second (aka TPS).

    2.) if the product is not able do resume SSL session (by caching SSL ID), just forget it!
    It won’t perform well and is not scalable at all.

    Note that having a load-balancer which is able to maintain affinity based on SSL session ID is really important. You can understand why now.

    3.) bear in mind that the asymmetric key size may have a huge impact on performance.
    Of course, the bigger the asymmetric key size is, the harder it will be for an attacker to break the generated symmetric key.

    4.) stud is young, but seems promising
    By the way, stud has included HAProxy Technologies patches from @emericbr, so if you use a recent stud version, you may have the same result as us.

    5.) euh, let’s read again the results… If we consider that your user would renegociate every 100 request and that the average object size you want to encrypt is 4K, you could get 2300 SSL transaction per second on a small Intel Atom @1.66GHZ!!!!
    Imagine what you could do with a dual CPU core i7!!!

    By the way, we’re glad that the stud developers have integrated our patches into main stud branch:

    Related links

How to play with maxconn to avoid server slowness or crash

Using Aloha load balancer and HAProxy, it is easy to protect any application or web server against unexpected high load.

Introduction

The response time of web servers is directly related to the number of requests they have to manage at the same time. And the response time is not linearly linked to the number of requests, it looks like exponential.
The graph below shows a server response time compared to the number of simultaneous users browsing the website:

Simultaneous connections limiting

Simultaneous connections limiting is basically a number (aka the limit) a load balancer will consider as the maximum number of requests to send to a backend server at the same time.
Of course, since HAProxy has such a function, Aloha load-balancer does.

Smart handling of requests peak with HAProxy

The meaning is too prevent too many requests to be forwarded to an application server, by adding a limit for simultaneous requests for each server of the backend.

Fortunately, HAProxy would not reject any request over the limit, unlike some other load balancer does.

HAProxy use a queueing system and will wait for the backend server to be able to answer. This mechanism will add slow delays to request in the queue, but it has a few advantages :

  • no client request are rejected
  • every request can be faster served than with an overloaded backend server
  • the delay is still acceptable (a few ms in queue)
  • your server won’t crash because of the spike

simultaneous requests limiting occurs on the server side: HAProxy will limit the number of concurrent request to the server despite what happens on the client side.
HAProxy will never refuse any client connection until the underlying server runs out of capacity.

Concrete numbers

If you read carefully the graph above, you can easily see that the more your server has to process requests at the same time, the longer each request will take to process.
The table below summarize the time spent by our example server to process 250 requests with different simultaneous requests limiting value:

Number of requests Simultaneous requests limit Average time per request Longuest response time in ms
250 10 9 225
250 20 9 112
250 30 9 75
250 50 25 125
250 100 100 250
250 150 225 305
250 250 625 625

It’s up to the website owner to know what will be the best limit to setup on HAProxy.
You can approximate it by using HTTP benchmark tools and by comparing average response time to constant number of request you send to your backend server.

From the example above, we can see we would get the best of this backend server by setting up the limit to 30.
Setting up a limit too low would implies queueing request for a longer time and setting it too high would be counter-productive by slowing down each request because of server capacity.

HAProxy simultaneous requests limiting configuration

The simultaneous requests limiting configuration is made with the maxconn keyword on the server line definition.
Example:

frontend APPLI1
	bind :80
	mode http
	option http-server-close
	default_backend APPLI1

backend APPLI1
	balance roundrobin
	mode http
	server server1 srv1:80 maxconn 30
	server server2 srv2:80 maxconn 30

Links