DISKD refers to some features in Squid-2.4 and later to improve Disk I/O performance. The basic idea is that each cache_dir has its own diskd child process. The diskd process performs all disk I/O operations (open, close, read, write, unlink) for the cache_dir. Message queues are used to send requests and responses between the Squid and diskd processes. Shared memory is used for chunks of data to be read and written.
Yes. We benchmarked Squid-2.4 with DISKD at the Second IRCache Bake-Off. The results are also described here. At the bakeoff, we got 160 req/sec with diskd. Without diskd, we'd have gotten about 40 req/sec.
You need to run Squid version 2.4 or later. Your operating system must support message queues, and shared memory.
To configure Squid for DISKD, use the --enable-storeio option:
% ./configure --enable-storeio=diskd,ufs
You didn't put diskd in the list of storeio modules as described above. You need to run configure and and recompile Squid.
No. Diskd uses the same storage scheme as the standard "UFS" type. It only changes how I/O is performed.
Most Unix operating systems have message queue support by default. One way to check is to see if you have an ipcs command.
However, you will likely need to increase the message queue parameters for Squid. Message queue implementations normally have the following parameters:
Maximum number of bytes per message queue.
Maximum number of message queue identifiers (system wide).
Maximum number of message segments per queue.
Size of a message segment.
Maximum number of messages (system wide).
Maximum size of a whole message. On some systems you may need to increase this limit. On other systems, you may not be able to change it.
The messages between Squid and diskd are 32 bytes for 32-bit CPUs and 40 bytes for 64-bit CPUs. Thus, MSGSSZ should be 32 or greater. You may want to set it to a larger value, just to be safe.
We'll have two queues for each cache_dir -- one in each direction. So, MSGMNI needs to be at least two times the number of cache_dir's.
I've found that 75 messages per queue is about the limit of decent performance. If each diskd message consists of just one segment (depending on your value of MSGSSZ), then MSGSEG should be greater than 75.
MSGMNB and MSGTQL affect how many messages can be in the queues at one time. Diskd messages shouldn't be more than 40 bytes, but let's use 64 bytes to be safe. MSGMNB should be at least 64*75. I recommend rounding up to the nearest power of two, or 8192.
MSGTQL should be at least 75 times the number of cache_dir's that you'll have.
Your kernel must have
options SYSVMSG
You can set the parameters in the kernel as follows. This is just an example. Make sure the values are appropriate for your system:
options MSGMNB=8192 # max # of bytes in a queue options MSGMNI=40 # number of message queue identifiers options MSGSEG=512 # number of message segments per queue options MSGSSZ=64 # size of a message segment options MSGTQL=2048 # max messages in system
You can set the parameters in the kernel as follows. This is just an example. Make sure the values are appropriate for your system:
option MSGMNB=16384 # max characters per message queue option MSGMNI=40 # max number of message queue identifiers option MSGSEG=2048 # max number of message segments in the system option MSGSSZ=64 # size of a message segment (Must be 2^N) option MSGTQL=1024 # max amount of messages in the system
Message queue support seems to be in the kernel by default. Setting the options is as follows:
options MSGMNB="8192" # max # bytes on queue options MSGMNI="40" # # of message queue identifiers options MSGMAX="2048" # max message size options MSGTQL="2048" # # of system message headers
If you have a newer version (DU64), then you can probably use sysconfig instead. To see what the current IPC settings are run
# sysconfig -q ipcTo change them make a file like this called ipc.stanza:
ipc:
msg-max = 2048
msg-mni = 40
msg-tql = 2048
msg-mnb = 8192
then run
# sysconfigdb -a -f ipc.stanzaYou have to reboot for the change to take effect.
Stefan Köpsell reports that if you compile sysctl support into your kernel, then you can change the following values:
Winfried Truemper reports: The default values should be large enough for most common cases. You can modify the message queue configuration by writing to these files:
Refer to Demangling Message Queues in Sunworld Magazine.
I don't think the above article really tells you how to set the parameters. You do it in /etc/system with lines like this:
set msgsys:msginfo_msgmax=2048 set msgsys:msginfo_msgmnb=8192 set msgsys:msginfo_msgmni=40 set msgsys:msginfo_msgssz=64 set msgsys:msginfo_msgtql=2048
Of course, you must reboot whenever you modify /etc/system before changes take effect.
Shared memory uses a set of parameters similar to the ones for message queues. The Squid DISKD implementation uses one shared memory area for each cache_dir. Each shared memory area is about 800 kilobytes in size. You may need to modify your system's shared memory parameters:
Maximum number of shared memory segments per process.
Maximum number of shared memory segments for the whole system.
Largest shared memory segment size allowed.
Total amount of shared memory that can be used.
For Squid and DISKD, SHMMNI and SHMMNI must be greater than or equal to the number of cache_dir's that you have. SHMMAX must be at least 800 kilobytes. SHMALL must be at least SHMMAX 800 kilobytes multiplied by the number of cache_dir's.
Your kernel must have
options SYSVSHM
You can set the parameters in the kernel as follows. This is just an example. Make sure the values are appropriate for your system:
options SHMSEG=16 # max shared mem id's per process options SHMMNI=32 # max shared mem id's per system options SHMMAX=2097152 # max shared memory segment size (bytes) options SHMALL=4096 # max amount of shared memory (pages)
OpenBSD is similar to FreeBSD, except you must use option instead of options, and SHMMAX is in pages instead of bytes:
option SHMSEG=16 # max shared mem id's per process option SHMMNI=32 # max shared mem id's per system option SHMMAX=2048 # max shared memory segment size (pages) option SHMALL=4096 # max amount of shared memory (pages)
Message queue support seems to be in the kernel by default. Setting the options is as follows:
options SHMSEG="16" # max shared mem id's per process options SHMMNI="32" # max shared mem id's per system options SHMMAX="2097152" # max shared memory segment size (bytes) options SHMALL=4096 # max amount of shared memory (pages)
If you have a newer version (DU64), then you can probably use sysconfig instead. To see what the current IPC settings are run
# sysconfig -q ipcTo change them make a file like this called ipc.stanza:
ipc:
shm-seg = 16
shm-mni = 32
shm-max = 2097152
shm-all = 4096
then run
# sysconfigdb -a -f ipc.stanzaYou have to reboot for the change to take effect.
Winfried Truemper reports: The default values should be large enough for most common cases. You can modify the shared memory configuration by writing to these files:
Stefan Köpsell reports that if you compile sysctl support into your kernel, then you can change the following values:
Refer to Shared memory uncovered in Sunworld Magazine.
To set the values, you can put these lines in /etc/system:
set shmsys:shminfo_shmmax=2097152 set shmsys:shminfo_shmmni=32 set shmsys:shminfo_shmseg=16
Yes, this is a little problem sometimes. Seems like the operating system gets confused and doesn't always release shared memory and message queue resources when processes exit, especially if they exit abnormally. To fix it you can ``manually'' clear the resources with the ipcs command. Add this command into your RunCache or squid_start script:
ipcs | grep '^[mq]' | awk '{printf "ipcrm -%s %s\n", $1, $2}' | /bin/sh
In the source code, these are called magic1 and magic2. These numbers refer to the number of oustanding requests on a message queue. They are specified on the cache_dir option line, after the L1 and L2 directories:
cache_dir diskd /cache1 1024 16 256 Q1=72 Q2=64
If there are more than Q1 messages outstanding, then Squid will intentionally fail to open disk files for reading and writing. This is a load-shedding mechanism. If your cache gets really really busy and the disks can not keep up, Squid bypasses the disks until the load goes down again.
If there are more than Q2 messages outstanding, then the main Squid process ``blocks'' for a little bit until the diskd process services some of the messages and sends back some replies.
Reasonable Q1 and Q2 values are 64 and 72. If you would rather have good hit ratio and bad response time, set Q1 > Q2. Otherwise, if you would rather have good response time and bad hit ratio, set Q1 < Q2.