Very high load on P4 machines with 2.4.28
Hello,
We have several machines with similar configurations
0000:00:00.0 Host bridge: Intel Corp. 82875P Memory Controller Hub (rev 02) 0000:00:01.0 PCI bridge: Intel Corp. 82875P Processor to AGP Controller (rev 02) 0000:00:1e.0 PCI bridge: Intel Corp. 82801 PCI Bridge (rev c2)
0000:00:1f.0 ISA bridge: Intel Corp. 82801EB/ER (ICH5/ICH5R) LPC Bridge (rev 02) 0000:00:1f.2 IDE interface: Intel Corp. 82801EB (ICH5) Serial ATA 150 Storage Controller (rev 02) 0000:00:1f.3 SMBus: Intel Corp. 82801EB/ER (ICH5/ICH5R) SMBus Controller (rev 02) 0000:02:09.0 VGA compatible controller: ATI Technologies Inc Rage XL (rev 27) 0000:02:0a.0 Ethernet controller: Intel Corp. 82541EI Gigabit Ethernet Controller (Copper) 0000:02:0b.0 Ethernet controller: Intel Corp. 82541EI Gigabit Ethernet Controller (Copper)
and
0000:00:00.0 Host bridge: Intel Corp. 82845G/GL[Brookdale-G]/GE/PE DRAM Controller/Host-Hub Interface (rev 03) 0000:00:02.0 VGA compatible controller: Intel Corp. 82845G/GL[Brookdale-G]/GE Chipset Integrated Graphics Device (rev 03) 0000:00:1e.0 PCI bridge: Intel Corp. 82801 PCI Bridge (rev 82)
0000:00:1f.0 ISA bridge: Intel Corp. 82801DB/DBL (ICH4/ICH4-L) LPC Bridge (rev 02)
0000:00:1f.1 IDE interface: Intel Corp. 82801DB/DBL (ICH4/ICH4-L) UltraATA-100 IDE Controller (rev 02) 0000:00:1f.3 SMBus: Intel Corp. 82801DB/DBL/DBM (ICH4/ICH4-L/ICH4-M) SMBus Controller (rev 02) 0000:01:05.0 Ethernet controller: Intel Corp. 82540EM Gigabit Ethernet Controller (rev 02) 0000:01:06.0 Ethernet controller: Intel Corp. 82540EM Gigabit Ethernet Controller (rev 02)
equipped with 2.6Ghz P4 CPUs, 1Gb of ram, 2-4gb of swap, the kernel config is attached. The machines have normal load averages hovering not higher than 7.0, depending on the time of the day etc. Two of the machines run 2.4.25, one 2.4.27 and they work fine. When booted with 2.4.28, though (compiled with Debian's gcc 2.3.5, with p3 or p4 CPU selected in the config), the load is climbing very fast and hovers around a value 3-4 times higher than with the older kernels. Booted back in the old kernel, the load comes to its usual level. The logs suggest nothing, no errors, nothing unusual is happening.
Has anyone had similar problems with 2.4.28 in an environment resembling the above? Could it be a problem with highmem i/o?
tia,
marek
Hi,
On Tue, Jan 04, 2005 at 08:56:36PM +0100, Marek Habersack wrote:
(...)
Quoted 10 lines
equipped with 2.6Ghz P4 CPUs, 1Gb of ram, 2-4gb of swap, the kernel config is attached. The machines have normal load averages hovering not higher than 7.0, depending on the time of the day etc. Two of the machines run 2.4.25, one 2.4.27 and they work fine. When booted with 2.4.28, though (compiled with Debian's gcc 2.3.5, with p3 or p4 CPU selected in the config), the load is climbing very fast and hovers around a value 3-4 times higher than with the older kernels. Booted back in the old kernel, the load comes to its usual level. The logs suggest nothing, no errors, nothing unusual is happening.
Quoted 2 lines
Has anyone had similar problems with 2.4.28 in an environment resembling the above? Could it be a problem with highmem i/o?
Never encountered yet ! Could you provide some indications about the type of work (I/O, network, CPU, scripts execution, #of processes, etc...) ?
Regards,
Willy
Oh, while I'm at it, are you using hyperthreading, and if so, could you disable it ? I have seen many cases where it degrades performances significantly (eg: highly loaded user space network applications).
Willy
Quoted 41 lines · Marek Habersack
Hello,
We have several machines with similar configurations
0000:00:00.0 Host bridge: Intel Corp. 82875P Memory Controller Hub (rev 02) 0000:00:01.0 PCI bridge: Intel Corp. 82875P Processor to AGP Controller (rev 02) 0000:00:1e.0 PCI bridge: Intel Corp. 82801 PCI Bridge (rev c2)
0000:00:1f.0 ISA bridge: Intel Corp. 82801EB/ER (ICH5/ICH5R) LPC Bridge (rev 02) 0000:00:1f.2 IDE interface: Intel Corp. 82801EB (ICH5) Serial ATA 150 Storage Controller (rev 02) 0000:00:1f.3 SMBus: Intel Corp. 82801EB/ER (ICH5/ICH5R) SMBus Controller (rev 02) 0000:02:09.0 VGA compatible controller: ATI Technologies Inc Rage XL (rev 27) 0000:02:0a.0 Ethernet controller: Intel Corp. 82541EI Gigabit Ethernet Controller (Copper) 0000:02:0b.0 Ethernet controller: Intel Corp. 82541EI Gigabit Ethernet Controller (Copper)and
0000:00:00.0 Host bridge: Intel Corp. 82845G/GL[Brookdale-G]/GE/PE DRAM Controller/Host-Hub Interface (rev 03) 0000:00:02.0 VGA compatible controller: Intel Corp. 82845G/GL[Brookdale-G]/GE Chipset Integrated Graphics Device (rev 03) 0000:00:1e.0 PCI bridge: Intel Corp. 82801 PCI Bridge (rev 82)
0000:00:1f.0 ISA bridge: Intel Corp. 82801DB/DBL (ICH4/ICH4-L) LPC Bridge (rev 02)
0000:00:1f.1 IDE interface: Intel Corp. 82801DB/DBL (ICH4/ICH4-L) UltraATA-100 IDE Controller (rev 02) 0000:00:1f.3 SMBus: Intel Corp. 82801DB/DBL/DBM (ICH4/ICH4-L/ICH4-M) SMBus Controller (rev 02) 0000:01:05.0 Ethernet controller: Intel Corp. 82540EM Gigabit Ethernet Controller (rev 02) 0000:01:06.0 Ethernet controller: Intel Corp. 82540EM Gigabit Ethernet Controller (rev 02)equipped with 2.6Ghz P4 CPUs, 1Gb of ram, 2-4gb of swap, the kernel config is attached. The machines have normal load averages hovering not higher than 7.0, depending on the time of the day etc. Two of the machines run 2.4.25, one 2.4.27 and they work fine. When booted with 2.4.28, though (compiled with Debian's gcc 2.3.5, with p3 or p4 CPU selected in the config), the load is climbing very fast and hovers around a value 3-4 times higher than with the older kernels. Booted back in the old kernel, the load comes to its usual level. The logs suggest nothing, no errors, nothing unusual is happening.
Has anyone had similar problems with 2.4.28 in an environment resembling the above? Could it be a problem with highmem i/o?
tia,
marek
On Tue, Jan 04, 2005 at 11:03:13PM +0100, Willy Tarreau scribbled:
Quoted 14 lines
Hi,
On Tue, Jan 04, 2005 at 08:56:36PM +0100, Marek Habersack wrote:
(...)Quoted 10 lines
equipped with 2.6Ghz P4 CPUs, 1Gb of ram, 2-4gb of swap, the kernel config is attached. The machines have normal load averages hovering not higher than 7.0, depending on the time of the day etc. Two of the machines run 2.4.25, one 2.4.27 and they work fine. When booted with 2.4.28, though (compiled with Debian's gcc 2.3.5, with p3 or p4 CPU selected in the config), the load is climbing very fast and hovers around a value 3-4 times higher than with the older kernels. Booted back in the old kernel, the load comes to its usual level. The logs suggest nothing, no errors, nothing unusual is happening.
Quoted 5 lines
Quoted 2 lines
Has anyone had similar problems with 2.4.28 in an environment resembling the above? Could it be a problem with highmem i/o?
Never encountered yet ! Could you provide some indications about the type of work (I/O, network, CPU, scripts execution, #of processes, etc...) ?
Of course. Here's some information:
the machines (with exception of one) are virtual hosting servers running apache with a lot of customer-provided perl scripts, php code, mysql (quite heavily used). The bandwidth used ranges from 4Mbit/s (the non-virtual box) to 24Mbit/s. The number of processes ranges from ~300 to ~600. Interestingly enough, the machine with the highest load average is the one generating 4Mbit/s and the one with 24Mbit/s has the smallest load average value. The latter also suffers from the biggest loadavg increase. All of the virtual machines have iptables accounting chains for each configured IP (there are between 62 IP numbers on one and 32 on the other). The virtual boxes have two 80GB SATA drives raided with softraid. The non-virtual box has a single IDE drive, no raid.
Some diagnostics (virtuals running 2.4.25 with ow1, non-virtual 2.4.27 with ow1):
(virtual #1)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 0 0 108364 33376 3032 382452 1 1 85 107 79 76 42 7 52 0
# iostat
avg-cpu: %user %nice %sys %iowait %idle
41.86 0.00 6.63 0.00 51.51Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn dev8-0 44.33 91.04 761.64 32859202 274901552 dev8-1 44.07 88.28 759.27 31863034 274048056
# cat /proc/interrupts
CPU0
0: 36164314 XT-PIC timer
1: 2 XT-PIC keyboard
2: 0 XT-PIC cascade
4: 5 XT-PIC serial
8: 4 XT-PIC rtc
10: 304927622 XT-PIC eth0
12: 786373 XT-PIC eth1
14: 31209236 XT-PIC libata
15: 1 XT-PIC ide1
NMI: 0
ERR: 0(virtual #2, the 24Mbit/s one)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 5 3 172448 13084 1208 304048 4 4 90 50 109 117 19 8 73 0
# iostat
avg-cpu: %user %nice %sys %iowait %idle
18.76 0.00 7.98 0.00 73.26Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn
dev8-0 28.66 465.98 323.12 168199261 116634204 dev8-1 28.26 458.50 314.57 165501660 113547124
# cat /proc/interrupts
CPU0
0: 36164279 XT-PIC timer
1: 2 XT-PIC keyboard
2: 0 XT-PIC cascade
4: 5 XT-PIC serial
8: 4 XT-PIC rtc
10: 713628758 XT-PIC eth0
12: 1452211 XT-PIC eth1
14: 20094643 XT-PIC libata
15: 1 XT-PIC ide1
NMI: 0
ERR: 0(the non-virtual)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 60 0 70300 115960 0 369244 0 0 79 32 90 45 73 7 21 0
# iostat
avg-cpu: %user %nice %sys %iowait %idle
72.64 0.03 6.63 0.00 20.71Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn dev3-0 7.62 54.77 91.76 19834569 33227488
# cat /proc/interrupts
CPU0
0: 36284746 XT-PIC timer
1: 2 XT-PIC keyboard
2: 0 XT-PIC cascade
4: 6 XT-PIC serial
8: 4 XT-PIC rtc
10: 787346 XT-PIC eth1
11: 120939540 XT-PIC eth0
14: 2743009 XT-PIC ide0
15: 1069794 XT-PIC ide1
NMI: 0
ERR: 0sar doesn't show anything interesting for any of the boxes. We tried to run 2.6.10 on the non-virtual box to see whether it could cope with its load problems, it crashed after one day with no trace in the logs as to what might have caused the problem (this box runs some software that's pretty bursty and CPU-intensive giving sometimes around 50-80 apache processes running at a time). One other interesting thing to note is that we have one other box with the similar configuration to the virtuals (also a virtual host) but it runs 2.4.28 with SMP+HT enabled - no load problems there at all. In the past we had a problem with P4 machines running very slowly when the kernel was compiled with the P4 CPU target, but it was on different hardware (an MSI mobo, the boxes above all have Supermicro mobos) and compiling the kernel with P3 CPU target helped with the slowness. That's why we tried the same "trick" here - but this time it didn't help. Let me know if you need more info,
thanks for your help, best regards
marek
On Tue, Jan 04, 2005 at 11:05:21PM +0100, Willy Tarreau scribbled:
Quoted 1 line
Oh, while I'm at it, are you using hyperthreading, and if so, could you
yes, as I wrote in the mail I've just sent - on one box which does NOT exhibit the problem... :)
Quoted 2 lines
disable it ? I have seen many cases where it degrades performances significantly (eg: highly loaded user space network applications).
We saw it in two cases as well, that's why in general we don't run with HT enabled (although we do test the boxes with it and if it behaves, we leave it on).
best regards,
marek
Hi,
Quoted 3 lines · Marek Habersack
Interestingly enough, the machine with the highest load average is the one generating 4Mbit/s and the one with 24Mbit/s has the smallest load average value.
This is common with multi-process servers like apache if the link is saturated, because data takes more time to reach the client, so you have a higher concurrency.
Quoted 11 lines
The latter also suffers from the biggest loadavg increase.
All of the virtual machines have iptables accounting chains for each configured IP (there are between 62 IP numbers on one and 32 on the other). The virtual boxes have two 80GB SATA drives raided with softraid. The non-virtual box has a single IDE drive, no raid.(virtual #2, the 24Mbit/s one)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 5 3 172448 13084 1208 304048 4 4 90 50 109 117 19 8 73 0
I don't like something : with 73% idle, you have 5 processes in the rq. I think this machine writes logs synchronously to disks, or stores SSL sessions on a real disk and waits for writes. A tmpfs would be a great help.
You can try to trace the processes activity with :
# strace -Te write <process pid> It will display the time elapsed in each write() syscall, you'll find the fds in /proc/<pid>/fd. You may notice big times on logs or ssl sessions.
Quoted 5 lines
(the non-virtual)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 60 0 70300 115960 0 369244 0 0 79 32 90 45 73 7 21 0
Same note for this one, although it does more user space work (php? ssl?). It's possible that some change in 2.4.28 touches the I/O subsystem and increases your wait I/O time in this particular application. (...)
Quoted 4 lines
One other interesting thing to note is that we have one other box with the similar configuration to the virtuals (also a virtual host) but it runs 2.4.28 with SMP+HT enabled - no load problems there at all.
So, to contradict myself, have you tried enabling HT on other boxes which suffer from the load ?
Quoted 1 line
Let me know if you need more info,
You have send fairly enough info right now. Other than I/O work, I have no idea. You may want to play with /proc/sys/vm/{bdflush,max-readahead} and others to see if it changes things.
If your load is bursty, it might help to reduce the ratio of dirty blocks before flushing (first field in bdflush), because although writes will start more often, they will take fewer time.
I already have solved similar problems by disabling keep-alive to decrease the number of processes.
Regards,
Willy
On Wed, Jan 05, 2005 at 06:28:41AM +0100, Willy Tarreau scribbled:
Quoted 1 line
Hi,
Hello,
Quoted 8 lines · Marek Habersack
Quoted 3 lines
Interestingly enough, the machine with the highest load average is the one generating 4Mbit/s and the one with 24Mbit/s has the smallest load average value.
This is common with multi-process servers like apache if the link is saturated, because data takes more time to reach the client, so you have a higher concurrency.
The link isn't saturated - we have a 200Mbit/s margin atm. It's not a bandwidth problem, that's certain.
Quoted 6 lines
Quoted 5 lines
The latter also suffers from the biggest loadavg increase.
All of the virtual machines have iptables accounting chains for each configured IP (there are between 62 IP numbers on one and 32 on the other). The virtual boxes have two 80GB SATA drives raided with softraid. The non-virtual box has a single IDE drive, no raid.
Quoted 8 lines
Quoted 5 lines
(virtual #2, the 24Mbit/s one)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 5 3 172448 13084 1208 304048 4 4 90 50 109 117 19 8 73 0I don't like something : with 73% idle, you have 5 processes in the rq. I think this machine writes logs synchronously to disks, or stores SSL sessions on a
the only synchronously written logs are auth.log and mail.err, SSL is there indeed, but the site is hardly ever accessed (as of a while ago, the box has a load of 0.75, pushing out 14Mbit/s. With 2.4.28 last week it was around 10.0 in the same conditions).
Quoted 1 line
real disk and waits for writes. A tmpfs would be a great help.
The only thigs writing to disk on regular basis (except for syslog and apache for logs) are the php session files, one tdb database for traffic data and mysql (which might be using fsync - can that be the cause of the i/o slowness?). But, in any case, the machine behaves well under kernels other than 2.4.28.
Quoted 5 lines
You can try to trace the processes activity with :
# strace -Te write <process pid> It will display the time elapsed in each write() syscall, you'll find the fds in /proc/<pid>/fd. You may notice big times on logs or ssl sessions.
nope... the times are in the range 0.000008 to 0.000045...
Quoted 7 lines
Quoted 5 lines
(the non-virtual)
# vmstat procs -----------memory---------- ---swap-- -----io---- --system------cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 60 0 70300 115960 0 369244 0 0 79 32 90 45 73 7 21 0Same note for this one, although it does more user space work (php? ssl?).
poorly written perl scripts
Quoted 3 lines
It's possible that some change in 2.4.28 touches the I/O subsystem and increases your wait I/O time in this particular application. (...)
Any clues as to where too look? I examined the 2.4.28 changelog and saw nothing that would suggest such change, but then I'm not a kernel hacker, I might have easily missed something important.
Quoted 7 lines
Quoted 4 lines
One other interesting thing to note is that we have one other box with the similar configuration to the virtuals (also a virtual host) but it runs 2.4.28 with SMP+HT enabled - no load problems there at all.
So, to contradict myself, have you tried enabling HT on other boxes which suffer from the load ?
Yep, only one box boots fine with HT enabled (out of the ones with problems), the others just freeze (we thought it could have been the machine BIOS, but updating it didn't help)
Quoted 5 lines
Quoted 1 line
Let me know if you need more info,
You have send fairly enough info right now. Other than I/O work, I have no idea. You may want to play with
/proc/sys/vm/{bdflush,max-readahead}and others to see if it changes things.
At this point I think we're gonna run them under the older kernels and wait for 2.4.29 to see whether the problem still exists there. If it does, we'll try 2.6 on the machines and if that doesn't help, we'll do some more testing with 2.4.28 - we have our hands tied, since they are production machines and we cannot let them run with such degraded performance for too long...
Quoted 3 lines
If your load is bursty, it might help to reduce the ratio of dirty blocks before flushing (first field in bdflush), because although writes will start more often, they will take fewer time.
what about nfract_sync? Does it make sense to make it smaller as well? I've also decreased age_buffer to 15s
Quoted 2 lines
I already have solved similar problems by disabling keep-alive to decrease the number of processes.
Disabling keep-alive is a routine here... :) But, that is unlikely to be the cause since it's evidently a kernel thing.
Well, I'll see what good the bdflush changes do to the machines when they run under the "good" kernel and we'll schedule for some testing with 2.4.28 at some point.
thanks for your help, it's greately appreciated!
best regards,
marek
Quoted 35 lines · Marek Habersack
Hello,
We have several machines with similar configurations
0000:00:00.0 Host bridge: Intel Corp. 82875P Memory Controller Hub (rev 02) 0000:00:01.0 PCI bridge: Intel Corp. 82875P Processor to AGP Controller (rev 02) 0000:00:1e.0 PCI bridge: Intel Corp. 82801 PCI Bridge (rev c2)
0000:00:1f.0 ISA bridge: Intel Corp. 82801EB/ER (ICH5/ICH5R) LPC Bridge (rev 02) 0000:00:1f.2 IDE interface: Intel Corp. 82801EB (ICH5) Serial ATA 150 Storage Controller (rev 02) 0000:00:1f.3 SMBus: Intel Corp. 82801EB/ER (ICH5/ICH5R) SMBus Controller (rev 02) 0000:02:09.0 VGA compatible controller: ATI Technologies Inc Rage XL (rev 27) 0000:02:0a.0 Ethernet controller: Intel Corp. 82541EI Gigabit Ethernet Controller (Copper) 0000:02:0b.0 Ethernet controller: Intel Corp. 82541EI Gigabit Ethernet Controller (Copper)and
0000:00:00.0 Host bridge: Intel Corp. 82845G/GL[Brookdale-G]/GE/PE DRAM Controller/Host-Hub Interface (rev 03) 0000:00:02.0 VGA compatible controller: Intel Corp. 82845G/GL[Brookdale-G]/GE Chipset Integrated Graphics Device (rev 03) 0000:00:1e.0 PCI bridge: Intel Corp. 82801 PCI Bridge (rev 82)
0000:00:1f.0 ISA bridge: Intel Corp. 82801DB/DBL (ICH4/ICH4-L) LPC Bridge (rev 02)
0000:00:1f.1 IDE interface: Intel Corp. 82801DB/DBL (ICH4/ICH4-L) UltraATA-100 IDE Controller (rev 02) 0000:00:1f.3 SMBus: Intel Corp. 82801DB/DBL/DBM (ICH4/ICH4-L/ICH4-M) SMBus Controller (rev 02) 0000:01:05.0 Ethernet controller: Intel Corp. 82540EM Gigabit Ethernet Controller (rev 02) 0000:01:06.0 Ethernet controller: Intel Corp. 82540EM Gigabit Ethernet Controller (rev 02)equipped with 2.6Ghz P4 CPUs, 1Gb of ram, 2-4gb of swap, the kernel config is attached. The machines have normal load averages hovering not higher than 7.0, depending on the time of the day etc. Two of the machines run 2.4.25, one 2.4.27 and they work fine. When booted with 2.4.28, though (compiled with Debian's gcc 2.3.5, with p3 or p4 CPU selected in the config), the load is climbing very fast and hovers around a value 3-4 times higher than with the older kernels. Booted back in the old kernel, the load comes to its usual level. The logs suggest nothing, no errors, nothing unusual is happening.
Quoted 2 lines
Has anyone had similar problems with 2.4.28 in an environment resembling the above? Could it be a problem with highmem i/o?
Nothing that I'm aware of should cause such increase in loadavg.
Marek, can you please try 2.4.28-pre1 ?
On Wed, Jan 05, 2005 at 07:42:36AM -0200, Marcelo Tosatti scribbled:
[snip]
Quoted 6 lines
Quoted 2 lines
Has anyone had similar problems with 2.4.28 in an environment resembling the above? Could it be a problem with highmem i/o?
Nothing that I'm aware of should cause such increase in loadavg.
Marek, can you please try 2.4.28-pre1 ?
I should be able to schedule that on Friday. Currently, we're running 2.4.28 on one of the machines (the non-virtual) but with mem=800M to exclude highmem. So far - no problems... We'll see how it goes on
best regards,
marek
Quoted 2 lines · Marek Habersack
equipped with 2.6Ghz P4 CPUs, 1Gb of ram, 2-4gb of swap, the kernel config is attached.
It isn't...
Quoted 8 lines
The machines have normal load averages hovering not higher than 7.0, depending on the time of the day etc. Two of the machines run 2.4.25, one 2.4.27 and they work fine. When booted with 2.4.28, though (compiled with Debian's gcc 2.3.5, with p3 or p4 CPU selected in the config), the load is climbing very fast and hovers around a value 3-4 times higher than with the older kernels. Booted back in the old kernel, the load comes to its usual level. The logs suggest nothing, no errors, nothing unusual is happening.
You may try each of 2.4.28-pre{1,2,3} kernels with identical .config and pinpoint when did it happen.
--
vda