In normal operation, the main concern is the overhead for encryption, decryption and authentication of the actual IPsec ( ESP and/or AH) data packets. Tunnel setup and rekeying occur so much less frequently than packet processing that, in general, their overheads are not worth worrying about.
At startup, however, tunnel setup overheads may be significant. If you reboot a gateway and it needs to establish many tunnels, expect some delay. This and other issues for large gateways are discussed below.
The University of Wales at Aberystwyth has done quite detailed speed tests and put their results on the web.
Davide Cerri's thesis (in Italian) includes performance results for FreeS/WAN and for TLS. He posted an English summary on the mailing list.
Steve Bellovin used one of AT&T Research's FreeS/WAN gateways as his data source for an analysis of the cache sizes required for key swapping in IPsec. Available as text or PDF slides for a talk on the topic.
See also the NAI work mentioned in the next section.
We can come up with a formula that roughly relates CPU speed to the rate of IPsec processing possible. It is far from exact, but should be usable as a first approximation.
An analysis of authentication overheads for high-speed networks, including some tests using FreeS/WAN, is on the NAI Labs site. In particular, see figure 3 in this PDF document. Their estimates of overheads, measured in Pentium II cycles per byte processed are:
|Linux IP stack alone||no||no||no||5|
|IPsec without crypto||yes||no||no||11|
|IPsec, authentication only||yes||SHA-1||no||24|
|IPsec with encryption||yes||yes||yes||not tested|
Overheads for IPsec with encryption were not tested in the NAI work, but Antoon Bosselaers' web page gives cost for his optimised Triple DES implementation as 928 Pentium cycles per block, or 116 per byte. Adding that to the 24 above, we get 140 cycles per byte for IPsec with encryption.
At 140 cycles per byte, a 140 MHz machine can handle a megabyte -- 8 megabits -- per second. Speeds for other machines will be proportional to this. To saturate a link with capacity C megabits per second, you need a machine running at C * 140/8 = C * 17.5 MHz.
However, that estimate is not precise. It ignores the differences between:
and does not account for some overheads you will almost certainly have:
so we suggest using C * 25 to get an estimate with a bit of a built-in safety factor.
This covers only IP and IPsec processing. If you have other loads on your gateway -- for example if it is also working as a firewall -- then you will need to add your own safety factor atop that.
This estimate matches empirical data reasonably well. For example, Metheringham's tests, described below, show a 733 topping out between 32 and 36 Mbit/second, pushing data as fast as it can down a 100 Mbit link. Our formula suggests you need at least an 800 to handle a fully loaded 32 Mbit link. The two results are consistent.
Some examples using this estimation method:
|Interface||Machine speed in MHz|
|Minimum IPSEC gateway||Minimum with other
|DSL||1||25 MHz||whatever you have||133, or better if you have it|
|cable modem||3||75 MHz|
|any link, light load||5||125 MHz||133||200+, almost any surplus machine|
|Ethernet||10||250 MHz||surplus 266 or 300||500+|
|fast link, moderate load||20||500 MHz||500||800+, any current off-the-shelf PC|
|T3 or E3||45||1125 MHz||1200||1500+|
|fast Ethernet||100||2500 MHz||// not feasible with 3DES in software on current machines //|
Such an estimate is far from exact, but should be usable as minimum requirement for planning. The key observations are:
To date (March 2002) we have had only one mailing list report of measurements with the patch applied. It indicates that, at least for the tested load on that user's network, AES roughly doubles IPsec throughput. If further testing confirms this, it may prove possible to saturate an OC3 link in software on a high-end box.
Also, some work is being done toward support of hardware IPsec acceleration which might extend the range of requirements FreeS/WAN could meet.
CPU speed may be the main issue for IPsec performance, but of course it isn't the only one.
The current FreeS/WAN kernel code is largely single-threaded. It is SMP safe, and will run just fine on a multiprocessor machine ( discussion), but the load within the kernel is not shared effectively. This means that, for example to saturate a T3 -- which needs about a 1200 MHz machine -- you cannot expect something like a dual 800 to do the job.
On the other hand, SMP machines do tend to share loads well so -- provided one CPU is fast enough for the IPsec work -- a multiprocessor machine may be ideal for a gateway with a mixed load.
FreeS/WAN allows a single gateway machine to build tunnels to many others. There may, however, be some problems for large numbers as indicated in this message from the mailing list:
Subject: Re: Maximum number of ipsec tunnels? Date: Tue, 18 Apr 2000 From: "John S. Denker" <firstname.lastname@example.org> Christopher Ferris wrote: >> What are the maximum number ipsec tunnels FreeS/WAN can handle?? Henry Spencer wrote: >There is no particular limit. Some of the setup procedures currently >scale poorly to large numbers of connections, but there are (clumsy) >workarounds for that now, and proper fixes are coming. 1) "Large" numbers means anything over 50 or so. I routinely run boxes with about 200 tunnels. Once you get more than 50 or so, you need to worry about several scalability issues: a) You need to put a "-" sign in syslogd.conf, and rotate the logs daily not weekly. b) Processor load per tunnel is small unless the tunnel is not up, in which case a new half-key gets generated every 90 seconds, which can add up if you've got a lot of down tunnels. c) There's other bits of lore you need when running a large number of tunnels. For instance, systematically keeping the .conf file free of conflicts requires tools that aren't shipped with the standard freeswan package. d) The pluto startup behavior is quadratic. With 200 tunnels, this eats up several minutes at every restart. I'm told fixes are coming soon. 2) Other than item (1b), the CPU load depends mainly on the size of the pipe attached, not on the number of tunnels.
It is worth noting that item (1b) applies only to repeated attempts to re-key a data connection (IPsec SA, Phase 2) over an established keying connection (ISAKMP SA, Phase 1). There are two ways to reduce this overhead using settings in ipsec.conf(5):
The overheads for establishing keying connections (ISAKMP SAs, Phase 1) are lower because for these Pluto does not perform expensive operations before receiving a reply from the peer.
A gateway that does a lot of rekeying -- many tunnels and/or low settings for tunnel lifetimes -- will also need a lot of random numbers from the random(4) driver.
Even a 486 can handle a T1 line, according to this mailing list message:
Subject: Re: linux-ipsec: IPSec Masquerade Date: Fri, 15 Jan 1999 11:13:22 -0500 From: Michael Richardson . . . A 486/66 has been clocked by Phil Karn to do 10Mb/s encryption.. that uses all the CPU, so half that to get some CPU, and you have 5Mb/s. 1/3 that for 3DES and you get 1.6Mb/s....
and a piece of mail from project technical lead Henry Spencer:
Oh yes, and a new timing point for Sandy's docs... A P60 -- yes, a 60MHz Pentium, talk about antiques -- running a host-to-host tunnel to another machine shows an FTP throughput (that is, end-to-end results with a real protocol) of slightly over 5Mbit/s either way. (The other machine is much faster, the network is 100Mbps, and the ether cards are good ones... so the P60 is pretty definitely the bottleneck.)
From the above, and from general user experience as reported on the list, it seems clear that a cheap surplus machine -- a reasonable 486, a minimal Pentium box, a Sparc 5, ... -- can easily handle a home office or a small company connection using any of:
If available, we suggest using a Pentium 133 or better. This should ensure that, even under maximum load, IPsec will use less than half the CPU cycles. You then have enough left for other things you may want on your gateway -- firewalling, web caching, DNS and such.
Here is some additional data from the mailing list.
Subject: FreeSWAN (specically KLIPS) performance measurements Date: Thu, 01 Feb 2001 From: Nigel Metheringham <Nigel.Metheringham@intechnology.co.uk> I've spent a happy morning attempting performance tests against KLIPS (this is due to me not being able to work out the CPU usage of KLIPS so resorting to the crude measurements of maximum throughput to give a baseline to work out loading of a box). Measurements were done using a set of 4 boxes arranged in a line, each connected to the next by 100Mbit duplex ethernet. The inner 2 had an ipsec tunnel between them (shared secret, but I was doing measurements when the tunnel was up and running - keying should not be an issue here). The outer pair of boxes were traffic generators or traffic sink. The crypt boxes are Compaq DL380s - Uniprocessor PIII/733 with 256K cache. They have 128M main memory. Nothing significant was running on the boxes other than freeswan. The kernel was a 2.2.19pre7 patched with freeswan and ext3. Without an ipsec tunnel in the chain (ie the 2 inner boxes just being 100BaseT routers), throughput (measured with ttcp) was between 10644 and 11320 KB/sec With an ipsec tunnel in place, throughput was between 3268 and 3402 KB/sec These measurements are for data pushed across a TCP link, so the traffic on the wire between the 2 ipsec boxes would have been higher than this.... vmstat (run during some other tests, so not affecting those figures) on the encrypting box shows approx 50% system & 50% idle CPU - which I don't believe at all. Interactive feel of the box was significantly sluggish. I also tried running the kernel profiler (see man readprofile) during test runs. A box doing primarily decrypt work showed basically nothing happening - I assume interrupts were off. A box doing encrypt work showed the following:- Ticks Function Load ~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~ 956 total 0.0010 532 des_encrypt2 0.1330 110 MD5Transform 0.0443 97 kmalloc 0.1880 39 des_encrypt3 0.1336 23 speedo_interrupt 0.0298 14 skb_copy_expand 0.0250 13 ipsec_tunnel_start_xmit 0.0009 13 Decode 0.1625 11 handle_IRQ_event 0.1019 11 .des_ncbc_encrypt_end 0.0229 10 speedo_start_xmit 0.0188 9 satoa 0.0225 8 kfree 0.0118 8 ip_fragment 0.0121 7 ultoa 0.0365 5 speedo_rx 0.0071 5 .des_encrypt2_end 5.0000 4 _stext 0.0140 4 ip_fw_check 0.0035 2 rj_match 0.0034 2 ipfw_output_check 0.0200 2 inet_addr_type 0.0156 2 eth_copy_and_sum 0.0139 2 dev_get 0.0294 2 addrtoa 0.0143 1 speedo_tx_buffer_gc 0.0024 1 speedo_refill_rx_buf 0.0022 1 restore_all 0.0667 1 number 0.0020 1 net_bh 0.0021 1 neigh_connected_output 0.0076 1 MD5Final 0.0083 1 kmem_cache_free 0.0016 1 kmem_cache_alloc 0.0022 1 __kfree_skb 0.0060 1 ipsec_rcv 0.0001 1 ip_rcv 0.0014 1 ip_options_fragment 0.0071 1 ip_local_deliver 0.0023 1 ipfw_forward_check 0.0139 1 ip_forward 0.0011 1 eth_header 0.0040 1 .des_encrypt3_end 0.0833 1 des_decrypt3 0.0034 1 csum_partial_copy_generic 0.0045 1 call_out_firewall 0.0125 Hope this data is helpful to someone... however the lack of visibility into the decrypt side makes things less clear
Another user reported some results for connections with and without IP compression:
Subject: [Users] Speed with compression Date: Fri, 29 Jun 2001 From: John McMonagle <email@example.com> Did a couple tests with compression using the new 1.91 freeswan. Running between 2 sites with cable modems. Both using approximately 130 mhz pentium. Transferred files with ncftp. Compressed file was a 6mb compressed installation file. Non compressed was 18mb /var/lib/rpm/packages.rpm Compressed vpn regular vpn Compress file 42.59 kBs 42.08 kBs regular file 110.84 kBs 41.66 kBs Load was about 0 either way. Ping times were very similar a bit above 9 ms. Compression looks attractive to me.Later in the same thread, project technical lead Henry Spencer added:
> is there a reason not to switch compression on? I have large gateway boxes > connecting 3 connections, one of them with a measly DS1 link... Run some timing tests with and without, with data and loads representative of what you expect in production. That's the definitive way to decide. If compression is a net loss, then obviously, leave it turned off. If it doesn't make much difference, leave it off for simplicity and hence robustness. If there's a substantial gain, by all means turn it on. If both ends support compression and can successfully negotiate a compressed connection (trivially true if both are FreeS/WAN 1.91), then the crucial question is CPU cycles. Compression has some overhead, so one question is whether *your* data compresses well enough to save you more CPU cycles (by reducing the volume of data going through CPU-intensive encryption/decryption) than it costs you. Last time I ran such tests on data that was reasonably compressible but not deliberately contrived to be so, this generally was not true -- compression cost extra CPU cycles -- so compression was worthwhile only if the link, not the CPU, was the bottleneck. However, that was before the slow-compression bug was fixed. I haven't had a chance to re-run those tests yet, but it sounds like I'd probably see a different result.The bug he refers to was a problem with the compression libraries that had us using C code, rather than assembler, for compression. It was fixed before 1.91.
If you want to measure the loads FreeS/WAN puts on a system, note that tools such as top or measurements such as load average are more-or-less useless for this. They are not designed to measure something that does most of its work inside the kernel.
Here is a message from FreeS/WAN kernel programmer Richard Guy Briggs on this:
> I have a batch of boxes doing Freeswan stuff. > I want to measure the CPU loading of the Freeswan tunnels, but am > having trouble seeing how I get some figures out... > > - Keying etc is in userspace so will show up on the per-process > and load average etc (ie pluto's load) Correct. > - KLIPS is in the kernel space, and does not show up in load average > I think also that the KLIPS per-packet processing stuff is running > as part of an interrupt handler so it does not show up in the > /proc/stat system_cpu or even idle_cpu figures It is not running in interrupt handler. It is in the bottom half. This is somewhere between user context (careful, this is not userspace!) and hardware interrupt context. > Is this correct, and is there any means of instrumenting how much the > cpu is being loaded - I don't like the idea of a system running out of > steam whilst still showing 100% idle CPU :-) vmstat seems to do a fairly good job, but use a running tally to get a good idea. A one-off call to vmstat gives different numbers than a running stat. To do this, put an interval on your vmstat command line.and another suggestion from the same thread:
Subject: Re: Measuring the CPU usage of Freeswan Date: Mon, 29 Jan 2001 From: Patrick Michael Kane <firstname.lastname@example.org> The only truly accurate way to accurately track FreeSWAN CPU usage is to use a CPU soaker. You run it on an unloaded system as a benchmark, then start up FreeSWAN and take the difference to determine how much FreeSWAN is eating. I believe someone has done this in the past, so you may find something in the FreeSWAN archives. If not, someone recently posted a URL to a CPU soaker benchmark tool on linux-kernel.