14. Системно-зависимые сверхестественности :)

14.1 Solaris

select()

select(3c) не поддерживает более чем 1024 файловых дескрипторов. Скрипт configure по умолчанию должен включить poll() для Solaris. poll() позволяет вам использовать больше файловых дескрипторов, 8192 или более.

Для старых версий Squid вы можете включить poll() вручную, изменив HAVE_POLL в файле include/autoconf.h или добавив -DUSE_POLL=1 к the DEFINES в src/Makefile.

malloc

В libmalloc.a есть утечка памяти. Squid-вый configure не использует -lmalloc на Solaris.

DNS lookups and nscd

от David J N Begley.

DNS lookups can be slow because of some mysterious thing called ncsd. You should edit /etc/nscd.conf and make it say:

enable-cache hosts no

Apparently nscd serializes DNS queries thus slowing everything down when an application (such as Squid) hits the resolver hard. You may notice something similar if you run a log processor executing many DNS resolver queries - the resolver starts to slow.. right.. down.. . . .

According to Andres Kroonmaa, users of Solaris starting from version 2.6 and up should NOT completely disable nscd daemon. nscd should be running and caching passwd and group files, although it is suggested to disable hosts caching as it may interfere with DNS lookups.

Several library calls rely on available free FILE descriptors FD < 256. Systems running without nscd may fail on such calls if first 256 files are all in use.

Since solaris 2.6 Sun has changed the way some system calls work and is using nscd daemon as a implementor of them. To communicate to nscd Solaris is using undocumented door calls. Basically nscd is used to reduce memory usage of user-space system libraries that use passwd and group files. Before 2.6 Solaris cached full passwd file in library memory on the first use but as this was considered to use up too much ram on large multiuser systems Sun has decided to move implementation of these calls out of libraries and to a single dedicated daemon.

Запросы DNS и /etc/nsswitch.conf

от Jason Armistead.

The /etc/nsswitch.conf file determines the order of searches for lookups (amongst other things). You might only have it set up to allow NIS and HOSTS files to work. You definitely want the "hosts:" line to include the word dns, e.g.:

hosts: nis dns [NOTFOUND=return] files

DNS lookups and NIS

by Chris Tilbury.

Our site cache is running on a Solaris 2.6 machine. We use NIS to distribute authentication and local hosts information around and in common with our multiuser systems, we run a slave NIS server on it to help the response of NIS queries.

We were seeing very high name-ip lookup times (avg ~2sec) and ip->name lookup times (avg ~8 sec), although there didn't seem to be that much of a problem with response times for valid sites until the cache was being placed under high load. Then, performance went down the toilet.

After some time, and a bit of detective work, we found the problem. On Solaris 2.6, if you have a local NIS server running (ypserv) and you have NIS in your /etc/nsswitch.conf hosts entry, then check the flags it is being started with. The 2.6 ypstart script checks to see if there is a resolv.conf file present when it starts ypserv. If there is, then it starts it with the -d option.

This has the same effect as putting the YP_INTERDOMAIN key in the hosts table -- namely, that failed NIS host lookups are tried against the DNS by the NIS server.

This is a bad thing(tm)! If NIS itself tries to resolve names using the DNS, then the requests are serialised through the NIS server, creating a bottleneck (This is the same basic problem that is seen with nscd). Thus, one failing or slow lookup can, if you have NIS before DNS in the service switch file (which is the most common setup), hold up every other lookup taking place.

If you're running in this kind of setup, then you will want to make sure that

1. ypserv doesn't start with the -d flag.

2. you don't have the YP_INTERDOMAIN key in the hosts table (find the B=-b line in the yp Makefile and change it to B=)

We changed these here, and saw our average lookup times drop by up to an order of magnitude (~150msec for name-ip queries and ~1.5sec for ip-name queries, the latter still so high, I suspect, because more of these fail and timeout since they are not made so often and the entries are frequently non-existent anyway).

Tuning

Solaris 2.x - tuning your TCP/IP stack and more by Jens-S. Vckler

disk write error: (28) No space left on device

You might get this error even if your disk is not full, and is not out of inodes. Check your syslog logs (/var/adm/messages, normally) for messages like either of these:

NOTICE: realloccg /proxy/cache: file system full

NOTICE: alloc: /proxy/cache: file system full

In a nutshell, the UFS filesystem used by Solaris can't cope with the workload squid presents to it very well. The filesystem will end up becoming highly fragmented, until it reaches a point where there are insufficient free blocks left to create files with, and only fragments available. At this point, you'll get this error and squid will revise its idea of how much space is actually available to it. You can do a "fsck -n raw_device" (no need to unmount, this checks in read only mode) to look at the fragmentation level of the filesystem. It will probably be quite high (>15%).

Sun предлагает два решения этой проблемы. Одно из них стоит денег, второе - бесплатное, но его результатом может стать снижение производительности (хотя Sun утвердяет, что этого происходить не должно, даже если учыитывать слишком случайную природу доступа squid к диску).

Первое решение - купить копию VxFS, файловой системы Veritas. This is an extent-based filesystem and it's capable of having online defragmentation performed on mounted filesystems. This costs money, however (VxFS is not very cheap!)

Второе решение состоит в изменении определенных параметров файловой системы UFS. Unmount your cache filesystems and use tunefs to change optimization to "space" and to reduce the "minfree" value to 3-5% (under Solaris 2.6 and higher, very large filesystems will almost certainly have a minfree of 2% already and you shouldn't increase this). You should be able to get fragmentation down to around 3% by doing this, with an accompanied increase in the amount of space available.

Thanks to Chris Tilbury.

Solaris X86 and IPFilter

by Jeff Madison

Important update regarding Squid running on Solaris x86. I have been working for several months to resolve what appeared to be a memory leak in squid when running on Solaris x86 regardless of the malloc that was used. I have made 2 discoveries that anyone running Squid on this platform may be interested in.

Number 1: There is not a memory leak in Squid even though after the system runs for some amount of time, this varies depending on the load the system is under, Top reports that there is very little memory free. True to the claims of the Sun engineer I spoke to this statistic from Top is incorrect. The odd thing is that you do begin to see performance suffer substantially as time goes on and the only way to correct the situation is to reboot the system. This leads me to discovery number 2.

Number 2: There is some type of resource problem, memory or other, with IPFilter on Solaris x86. I have not taken the time to investigate what the problem is because we no longer are using IPFilter. We have switched to a Alteon ACE 180 Gigabit switch which will do the trans-proxy for you. After moving the trans-proxy, redirection process out to the Alteon switch Squid has run for 3 days strait under a huge load with no problem what so ever. We currently have 2 boxes with 40 GB of cached objects on each box. This 40 GB was accumulated in the 3 days, from this you can see what type of load these boxes are under. Prior to this change we were never able to operate for more than 4 hours.

Because the problem appears to be with IPFilter I would guess that you would only run into this issue if you are trying to run Squid as a transparent proxy using IPFilter. That makes sense. If there is anyone with information that would indicate my finding are incorrect I am willing to investigate further.

Changing the directory lookup cache size

by Mike Batchelor

On Solaris, the kernel variable for the directory name lookup cache size is ncsize. In /etc/system, you might want to try

set ncsize = 8192

or even higher. The kernel variable ufs_inode - which is the size of the inode cache itself - scales with ncsize in Solaris 2.5.1 and later. Previous versions of Solaris required both to be adjusted independently, but now, it is not recommended to adjust ufs_inode directly on 2.5.1 and later.

You can set ncsize quite high, but at some point - dependent on the application - a too-large ncsize will increase the latency of lookups.

Defaults are:

Solaris 2.5.1 : (max_nprocs + 16 + maxusers) + 64

Solaris 2.6/Solaris 7 : 4 * (max_nprocs + maxusers) + 320

The priority_paging algorithm

by Mike Batchelor

Another new tuneable (actually a toggle) in Solaris 2.5.1, 2.6 or Solaris 7 is the priority_paging algorithm. This is actually a complete rewrite of the virtual memory system on Solaris. It will page out application data last, and filesystem pages first, if you turn it on (set priority_paging = 1 in /etc/system). As you may know, the Solaris buffer cache grows to fill available pages, and under the old VM system, applications could get paged out to make way for the buffer cache, which can lead to swap thrashing and degraded application performance. The new priority_paging helps keep application and shared library pages in memory, preventing the buffer cache from paging them out, until memory gets REALLY short. Solaris 2.5.1 requires patch 103640-25 or higher and Solaris 2.6 requires 105181-10 or higher to get priority_paging. Solaris 7 needs no patch, but all versions have it turned off by default.